OctoInformBeta

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

Corollary 4.2: Two scalar constraints from Gram determinants of orthogonal residuals

Pairwise bounds constrain centroid norm gaps.

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

Competition between Gaussian cells involves the components of their centroids perpendicular to a chosen centroid. The library states norm gap bounds for finite Euclidean configurations and two corresponding inequalities for triples of vectors.

Write γd\gamma_d for standard Gaussian measure on ℝd\mathbb R^d, the Euclidean space with its usual inner product. For a Gaussian partition 𝒜=(A1,…,Am)\mathcal A=(A_1,\ldots,A_m), the unnormalized centroid ziz_i of AiA_i is the vector in ℝd\mathbb R^d given by (1), the partition value C=F(𝒜)C=F(\mathcal A) is the scalar given there, and, when C>0C>0, the normalized radius rir_i of AiA_i is the nonnegative scalar given there: zi=∫Aixdγd(x),C=F(𝒜)=∑i=1m∥zi∥2,ri=∥zi∥C,B=322π.z_i=\int_{A_i}x\,d\gamma_d(x),\qquad C=F(\mathcal A)=\sum_{i=1}^m\|z_i\|^2,\qquad r_i=\frac{\|z_i\|}{\sqrt C},\qquad B=\frac{3}{2\sqrt{2\pi}}.(1)

For z,y∈ℝdz,y\in\mathbb R^d, the negative projection coefficient α(z,y)\alpha(z,y) is a scalar, and the orthogonal residual Rz(y)R_z(y) is the vector defined in (2): α(z,y)={−⟨z,y⟩/∥z∥2,z≠0,0,z=0,Rz(y)=y+α(z,y)z.\alpha(z,y)= \begin{cases} -\langle z,y\rangle/\|z\|^2,&z\ne0,\\ 0,&z=0, \end{cases} \qquad R_z(y)=y+\alpha(z,y)z.(2) The residual Rzi(zj)R_{z_i}(z_j) is unnormalized; for a partition of value C>0C>0, the normalized centroid is zi/Cz_i/\sqrt C and the normalized residual is wj=Rzi(zj)/Cw_j=R_{z_i}(z_j)/\sqrt C.

The Gram determinant D(x,y)D(x,y) of two vectors in a real inner product space is the scalar defined by (3): D(x,y)=∥x∥2∥y∥2−⟨x,y⟩2.D(x,y)=\|x\|^2\|y\|^2-\langle x,y\rangle^2.(3)

The Claim

Theorem 1. Let z1,…,zk∈ℝdz_1,\ldots,z_k\in\mathbb R^d have sum zero and satisfy ⟨za,zb⟩<0\langle z_a,z_b\rangle<0 for distinct indices a,ba,b. Choose four distinct indices i,j,ℓ,hi,j,\ell,h such that zi≠0z_i\ne0 and ∥zi∥≤∥zh∥≤min⁡{∥zj∥,∥zℓ∥}\|z_i\|\le\|z_h\|\le\min\{\|z_j\|,\|z_\ell\|\}. Put va=Rzi(za)v_a=R_{z_i}(z_a) and αa=α(zi,za)\alpha_a=\alpha(z_i,z_a), and assume the pairwise residual bounds (4) for every a,b≠ia,b\ne i: D(va,vb)≤(3π)2∥zi∥2(1+αa)2(1+αb)2.D(v_a,v_b)\le \left(\frac3\pi\right)^2\|z_i\|^2(1+\alpha_a)^2(1+\alpha_b)^2.(4) The centroid norm gap bounds are (5): (∥zh∥2−∥zi∥2)(∥zj∥2+∥zℓ∥2−∥zh∥2−2∥zi∥2)≤8(3π)2∥zi∥2,32(∥zh∥2−∥zi∥2)2≤10(3π)2∥zi∥2.\begin{aligned} (\|z_h\|^2-\|z_i\|^2) (\|z_j\|^2+\|z_\ell\|^2-\|z_h\|^2-2\|z_i\|^2) &\le 8\left(\frac3\pi\right)^2\|z_i\|^2,\\ \frac32(\|z_h\|^2-\|z_i\|^2)^2 &\le 10\left(\frac3\pi\right)^2\|z_i\|^2. \end{aligned}(5)✦

Let u,v,wu,v,w belong to a real inner product space, and let t,c,L,κ,a,b,qt,c,L,\kappa,a,b,q be real numbers with a,b,q≥0a,b,q\ge0, a+b+q≤1a+b+q\le1 and t2≤ct^2\le c. Suppose the norm and inner product conditions (6) hold: c−t2≤∥w∥2,L−c−2t2≤∥u∥2+∥v∥2−∥w∥2,⟨w,u⟩≤0,⟨w,v⟩≤0,−⟨w,u⟩−⟨w,v⟩≤∥w∥2,D(w,u)≤κ2t2(1+q)2(1+a)2,D(w,v)≤κ2t2(1+q)2(1+b)2.\begin{aligned} c-t^2&\le\|w\|^2,& L-c-2t^2&\le\|u\|^2+\|v\|^2-\|w\|^2,\\ \langle w,u\rangle&\le0,&\langle w,v\rangle&\le0,\\ -\langle w,u\rangle-\langle w,v\rangle&\le\|w\|^2,&&\\ D(w,u)&\le\kappa^2t^2(1+q)^2(1+a)^2,&&\\ D(w,v)&\le\kappa^2t^2(1+q)^2(1+b)^2.&& \end{aligned}(6) The mixed norm gap bound is (7): (c−t2)(L−c−2t2)≤8κ2t2.(c-t^2)(L-c-2t^2)\le8\kappa^2t^2.(7)✦

For the same scalar conditions, suppose instead that u,v,wu,v,w are nonzero and satisfy (8): c−t2≤min⁡{∥u∥2,∥v∥2,∥w∥2},⟨u,v⟩≤0,⟨u,w⟩≤0,⟨v,w⟩≤0,D(u,v)≤κ2t2(1+a)2(1+b)2,D(u,w)≤κ2t2(1+a)2(1+q)2,D(v,w)≤κ2t2(1+b)2(1+q)2.\begin{aligned} c-t^2&\le\min\{\|u\|^2,\|v\|^2,\|w\|^2\},\\ \langle u,v\rangle&\le0,\qquad \langle u,w\rangle\le0,\qquad\langle v,w\rangle\le0,\\ D(u,v)&\le\kappa^2t^2(1+a)^2(1+b)^2,\\ D(u,w)&\le\kappa^2t^2(1+a)^2(1+q)^2,\\ D(v,w)&\le\kappa^2t^2(1+b)^2(1+q)^2. \end{aligned}(8) The squared norm gap bound is (9): 32(c−t2)2≤10κ2t2.\frac32(c-t^2)^2\le10\kappa^2t^2.(9)✦

In the paper, the result reads: for an extremizing partition of value C>B2C>B^2 with m≥5m\ge5 and ordered normalized radii r1≥⋯≥rm=t>0r_1\ge\cdots\ge r_m=t>0, writing L=r12+r22L=r_1^2+r_2^2 and c=r32c=r_3^2, one has (c−t2)(L−c−2t2)≤8(3/π)2t2(c-t^2)(L-c-2t^2)\le8(3/\pi)^2t^2 and 32(c−t2)2≤10(3/π)2t2\frac32(c-t^2)^2\le10(3/\pi)^2t^2.

What Is Checked

The bound (5) is stated over finite zero-sum Euclidean configurations with strictly negative inner products between distinct vectors, the specified norm ordering and the residual bounds (4); the bounds (7) and (9) are stated over triples in a real inner product space under their respective norm, sign and Gram determinant hypotheses.