OctoInformBeta

OpenAI preprint, September 2026 · 5. The numerical obstruction

Section 5 (proof of Theorem 1.2): Nonnegative weights on an interval of length one give the bound 1/41/4

Variance decomposes in finite weighted mixtures.

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

The numerical obstruction involves finite weighted real data. Let SS be a finite set. A system of normalized real weights on SS is a family (wi)i∈S(w_i)_{i\in S} of real numbers satisfying (1). ∑i∈Swi=1.\sum_{i\in S}w_i=1.(1)

For real data (xi)i∈S(x_i)_{i\in S}, the weighted mean x¯\overline x and weighted variance Var⁡w(x)\operatorname{Var}_w(x) are the real numbers defined by (2). x¯=∑i∈Swixi,Var⁡w(x)=∑i∈Swi(xi−x¯)2.\overline x=\sum_{i\in S}w_ix_i, \qquad \operatorname{Var}_w(x)=\sum_{i\in S}w_i(x_i-\overline x)^2.(2)

The Claim

Theorem 1. Let SS be a finite set and let (wi)i∈S(w_i)_{i\in S} be normalized real weights. For real data (mi)i∈S(m_i)_{i\in S} and (qi)i∈S(q_i)_{i\in S}, the weighted mixture identity (3) holds: 2∑i∈Swiqi−m¯2=∑i∈Swi(2qi−mi2)+Var⁡w(m).2\sum_{i\in S}w_iq_i-\overline m^{\,2} = \sum_{i\in S}w_i(2q_i-m_i^2)+\operatorname{Var}_w(m).(3)✦ If the weights are nonnegative and real data (xi)i∈S(x_i)_{i\in S} lie in [a,a+1][a,a+1] for a real number aa, the weighted variance satisfies (4): Var⁡w(x)≤14.\operatorname{Var}_w(x)\leq\frac14.(4)✦

In the paper, the result reads: for the semistable factors GiG_i of a balanced sheaf PP, with positive ranks uiu_i and total rank rr, writing f*f_* for the direct image, HH for the polarization and c1c_1 for the first Chern class, put wi=ui/rw_i=u_i/r, mi=c1(f*Gi)/(uiH2)m_i=c_1(f_*G_i)/(u_iH^2) and m¯=∑iwimi\overline m=\sum_iw_im_i; then the numerical invariant vv satisfies v(f*P)=∑iwiv(f*Gi)+∑iwi(mi−m¯)2v(f_*P)=\sum_iw_iv(f_*G_i)+\sum_iw_i(m_i-\overline m)^2, and if the mim_i lie in an interval of length one, the last sum is at most 1/41/4.

What Is Checked

The mixture identity is stated over finite real data with real weights of sum one, and the variance bound additionally assumes nonnegative weights and data in [a,a+1][a,a+1].