OctoInformBeta

OpenAI preprint, September 2026 · §2. Recursive colors and finite averages

§2.3 (proof of Proposition 2.3): variance and displacement: A boundary constant for recursive averages without a sign condition on the reference correlation

Variance and displacement constrain finite correlations.

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 comparison is between correlations among parent colors and correlations involving their children, averaged over a finite subset of a group. Let GG be a group and let EE be a real inner product space. The closed unit ball BB of EE is the set in (1). B={v∈E:∥v∥≤1}.B=\{v\in E:\|v\|\leq 1\}.(1)✦ In the interval notation for recursive colors, XI(g)∈BX_I(g)\in B denotes the vector associated with g∈Gg\in G and a basic dyadic interval II. Interval composition is I⋅J=sI(J)I\cdot J=s_I(J), where sI:[0,1]→Is_I:[0,1]\to I is the increasing affine map.

For a finite family v=(vi)i=1Dv=(v_i)_{i=1}^D in EE, its finite arithmetic mean MD(v)M_D(v) is the vector defined in (2). The unit-ball mean MD(v)M_D(v) of a nonempty finite family v=(vi)i=1Dv=(v_i)_{i=1}^D in BB is the vector given by its finite arithmetic mean.✦ MD(v)={1D∑i=1Dvi,D>0,0,D=0.M_D(v)= \begin{cases} \displaystyle\frac{1}{D}\sum_{i=1}^D v_i,&D>0,\\ 0,&D=0. \end{cases}(2)✦ For a finite subset A⊆GA\subseteq G and a real-valued function uu on GG, the normalized finite average 𝔼Au\mathbb E_Au is the real number defined in (3). 𝔼Au={1|A|∑g∈Au(g),A≠⌀,0,A=⌀.\mathbb E_Au= \begin{cases} \displaystyle\frac{1}{|A|}\sum_{g\in A}u(g),&A\ne\varnothing,\\ 0,&A=\varnothing. \end{cases}(3) The finite inner product correlation CA(v,w)C_A(v,w) of functions v,w:G→Ev,w:G\to E is the real number in (4). CA(v,w)=𝔼A⟨v(g),w(g)⟩.C_A(v,w)=\mathbb E_A\langle v(g),w(g)\rangle.(4)✦ For real parameters LL and δ\delta and a natural number DD, the boundary constant c(L,δ,D)c(L,\delta,D) is the real number in (5). c(L,δ,D)=δ2/L2−4/D4(1−1/D).c(L,\delta,D)= \frac{\delta^2/L^2-4/D}{4(1-1/D)}.(5)✦

The Claim

Theorem 1.

  1. Let D≥1D\geq1, let L,δ>0L,\delta>0, and let f:B→Bf:B\to B satisfy the Lipschitz and uniform displacement conditions (6). ∥f(u)−f(v)∥≤L∥u−v∥,δ≤∥f(u)−u∥(u,v∈B).\|f(u)-f(v)\|\leq L\|u-v\|, \qquad \delta\leq\|f(u)-u\| \quad(u,v\in B).(6) For families x=(xi)i=1Dx=(x_i)_{i=1}^D and z=(zi)i=1Dz=(z_i)_{i=1}^D in BB with xi=f(zi)x_i=f(z_i), set m=MD(x)m=M_D(x); then (7) holds. δ2≤L2D∑i=1D∥zi−m∥2.\delta^2\leq\frac{L^2}{D}\sum_{i=1}^D\|z_i-m\|^2.(7)✦

  2. Let A⊆GA\subseteq G be finite and nonempty, let D≥1D\geq1, and let xi,yij:G→Ex_i,y_{ij}:G\to E satisfy ∥xi(g)∥≤1\|x_i(g)\|\leq1 and ∥yij(g)∥≤1\|y_{ij}(g)\|\leq1 for g∈Ag\in A and 1≤i,j≤D1\leq i,j\leq D. For families of functions xi,yij:G→Ex_i,y_{ij}:G\to E indexed by 1≤i,j≤D1\leq i,j\leq D, the mean functions m,zi:G→Em,z_i:G\to E are defined by m(g)=MD((xk(g))k=1D)m(g)=M_D((x_k(g))_{k=1}^D) and zi(g)=MD((yij(g))j=1D)z_i(g)=M_D((y_{ij}(g))_{j=1}^D). Let α,η∈ℝ\alpha,\eta\in\mathbb R, and assume the parent, sibling and mixed correlation conditions (8). |CA(xj,xk)−α|≤η(j≠k),|CA(yij,yik)−α|≤η(j≠k),|CA(yij,xk)−α|≤η(k≠i).\begin{aligned} |C_A(x_j,x_k)-\alpha|&\leq\eta &&(j\ne k),\\ |C_A(y_{ij},y_{ik})-\alpha|&\leq\eta &&(j\ne k),\\ |C_A(y_{ij},x_k)-\alpha|&\leq\eta &&(k\ne i). \end{aligned}(8) For each 1≤i≤D1\leq i\leq D, the finite variance estimate is (9). 𝔼A∥zi(g)−m(g)∥2≤4D+4(1−1D)η.\mathbb E_A\|z_i(g)-m(g)\|^2 \leq\frac{4}{D}+4\left(1-\frac1D\right)\eta.(9)✦

  3. Take the data of the second part with D≥2D\geq2 and with xi,yij:G→Bx_i,y_{ij}:G\to B. Let L,δ>0L,\delta>0 and let f:B→Bf:B\to B satisfy (6). Assume xi(g)=f(zi(g))x_i(g)=f(z_i(g)) for g∈Ag\in A and 1≤i≤D1\leq i\leq D. Then (10) holds. c(L,δ,D)≤η.c(L,\delta,D)\leq\eta.(10)✦

In the paper, the result reads: for g∈Ag\in A and parent intervals I1,…,IDI_1,\ldots,I_D, set m(g)=D−1∑kXIk(g)m(g)=D^{-1}\sum_kX_{I_k}(g) and zi(g)=D−1∑jXIi⋅Ij(g)z_i(g)=D^{-1}\sum_jX_{I_i\cdot I_j}(g), exact recursion gives XIi(g)=f(zi(g))X_{I_i}(g)=f(z_i(g)), the displacement and Lipschitz bounds imply δ2≤(L2/D)∑i∥zi(g)−m(g)∥2\delta^2\leq(L^2/D)\sum_i\|z_i(g)-m(g)\|^2, and, if the separated parent, sibling and mixed correlations are all within η\eta of the same α\alpha, unit-ball bounds on diagonal and nested terms give 𝔼A∥zi−m∥2≤4/D+4(1−1/D)η\mathbb E_A\|z_i-m\|^2\leq4/D+4(1-1/D)\eta.

What Is Checked

The claims are stated over finite arrays in a real inner product space, with norm bounds on AA for the variance estimate and a self-map of BB for the recursive boundary estimate. Finite arithmetic means of nonempty families in BB again belong to BB.✦