OctoInformBeta

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

§4: Residual scores and covariance determinants: Projection coefficients and residual inner products for a nonzero base vector

Removing the component along a centroid.

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

Comparison with a Gaussian cell singles out its centroid as a base direction. The residual scores describe the components of the competing centroids transverse to that direction.

Let ℝd\mathbb R^d have its usual Euclidean inner product and let γd\gamma_d be standard Gaussian measure on this space. The unnormalized centroid z(A)∈ℝdz(A)\in\mathbb R^d of a measurable cell A⊆ℝdA\subseteq\mathbb R^d is the vector integral in (1): z(A)=∫Axdγd(x).z(A)=\int_A x\,d\gamma_d(x).(1) For an optimal Gaussian partition (Aj)j=1m(A_j)_{j=1}^m, write zj=z(Aj)z_j=z(A_j) and C=∑j=1m∥zj∥2C=\sum_{j=1}^m\|z_j\|^2 for its value.

The Claim

Definition 1 (Negative projection coefficient and orthogonal residual). For vectors z,y∈ℝdz,y\in\mathbb R^d, the negative projection coefficient az(y)a_z(y) is the real scalar defined by (2), and the orthogonal residual Rz(y)∈ℝdR_z(y)\in\mathbb R^d is the vector defined by (3): az(y)={−⟨z,y⟩/∥z∥2,z≠0,0,z=0.a_z(y)=\begin{cases} -\langle z,y\rangle/\|z\|^2,&z\ne0,\\ 0,&z=0. \end{cases}(2)✦ Rz(y)=y+az(y)z.R_z(y)=y+a_z(y)z.(3)✦

These residuals are unnormalized; for an optimal partition of value C>0C>0, the corresponding normalized residual is Rz(y)/CR_z(y)/\sqrt C.

Theorem 1 (Residual identities). Let z∈ℝdz\in\mathbb R^d be nonzero and let y,v∈ℝdy,v\in\mathbb R^d. The residuals satisfy (4), (5), and (6): ⟨z,Rz(y)⟩=0.\langle z,R_z(y)\rangle=0.(4)✦ ⟨Rz(y),Rz(v)⟩=⟨y,v⟩−az(y)az(v)∥z∥2.\langle R_z(y),R_z(v)\rangle =\langle y,v\rangle-a_z(y)a_z(v)\|z\|^2.(5)✦ ∥Rz(y)∥2=∥y∥2−az(y)2∥z∥2.\|R_z(y)\|^2=\|y\|^2-a_z(y)^2\|z\|^2.(6)✦ For a finite collection z1,…,zk∈ℝdz_1,\ldots,z_k\in\mathbb R^d with ∑j=1kzj=0\sum_{j=1}^kz_j=0, and an index ii with zi≠0z_i\ne0, the coefficients satisfy (7): ∑1≤j≤kj≠iazi(zj)=1.\sum_{\substack{1\le j\le k\\j\ne i}}a_{z_i}(z_j)=1.(7)✦ For a finite family of vectors with zero sum and a chosen nonzero vector, the negative projection coefficients of the other vectors relative to the chosen vector sum to one.

In the paper, the result reads: for an optimal Gaussian partition of value C>B2C>B^2, where B=3/(22π)B=3/(2\sqrt{2\pi}), having the smallest number m≥5m\ge5 of active cells among optimal partitions, order rj=∥zj∥/Cr_j=\|z_j\|/\sqrt C as r1≥⋯≥rm=t>0r_1\ge\cdots\ge r_m=t>0 and put L=r12+r22L=r_1^2+r_2^2, c=r32c=r_3^2, e=zm/∥zm∥e=z_m/\|z_m\|, aj=−⟨zm,zj⟩/∥zm∥2a_j=-\langle z_m,z_j\rangle/\|z_m\|^2, and wj=zj/C+ajtew_j=z_j/\sqrt C+a_jte for j<mj<m; then aj>0a_j>0, ∑j<maj=1\sum_{j<m}a_j=1, wj⟂ew_j\perp e, and the Gaussian scores ⟨wj,g⟩\langle w_j,g\rangle for g∼γdg\sim\gamma_d are independent of ⟨e,g⟩\langle e,g\rangle, with Hjk=⟨wj,wk⟩H_{jk}=\langle w_j,w_k\rangle, hj=Hjjh_j=H_{jj}, and Djk=hjhk−Hjk2D_{jk}=\sqrt{h_jh_k-H_{jk}^2}.

What Is Checked

The statements concern vectors in finite-dimensional real Euclidean space, with a nonzero base vector for the residual identities and a finite collection summing to zero for the coefficient sum identity.