OctoInformBeta

OpenAI preprint, September 2026 · §3. Rank coordinates and limiting measures

Lemma 3.1: Participating vertices control rank labels and spatial enhancements under joint weak convergence

Cubic bounds for limiting edge mass.

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

In a planar exchange process, an edge contributes the label of its leaving endpoint, whereas a participating vertex is counted once regardless of its degree. The question is how these two concentrations compare in a time window after weak convergence.

A configuration PP of NN planar points is a map from {0,…,N−1}\{0,\ldots,N-1\} to ℝ2\mathbb R^2. A configuration PP of NN planar points is in general position when it is injective and the orientation determinant det⁡(Pj−Pi,Pl−Pi)\det(P_j-P_i,P_l-P_i) is nonzero for distinct indices.✦ A generic configuration PP of NN planar points is a configuration in general position that has pairwise distinct first coordinates and pairwise distinct slopes (Pj,2−Pi,2)/(Pj,1−Pi,1)(P_{j,2}-P_{i,2})/(P_{j,1}-P_{i,1}) for unordered pairs i<ji<j, and satisfies one further condition: for three distinct unordered pairs, the intersection of the first two open segments contains no point of the third closed segment.✦

A uniform exchange batch system SS on PP, with RR batches, rank kk and exchange distance aa, consists of strictly increasing slopes s0,…,sRs_0,\ldots,s_R and marked sets A0,…,ARA_0,\ldots,A_R of size kk. The sweep projection of p=(x,y)p=(x,y) at slope ss is y−sxy-sx.✦ Separation at stage bb requires every marked point to have smaller projection at sbs_b than every unmarked point, and exactly aa indices leave the marked set at each batch. The leaving sets LbL_b, entering sets GbG_b, exchange edges ℰ(S)\mathcal E(S) and participating vertices V(S)V(S) are defined by (1). Lb=Ab\Ab+1,Gb=Ab+1\Ab,ℰ(S)=⋃b=0R−1(Lb×Gb),V(S)={u:(u,v)∈ℰ(S)}∪{v:(u,v)∈ℰ(S)}.\begin{aligned} L_b&=A_b\setminus A_{b+1}, & G_b&=A_{b+1}\setminus A_b,\\ \mathcal E(S)&=\bigcup_{b=0}^{R-1}(L_b\times G_b), &V(S)&=\{u:(u,v)\in\mathcal E(S)\} \cup\{v:(u,v)\in\mathcal E(S)\}. \end{aligned}(1) A generic batch configuration consists of a generic planar configuration PP and a uniform exchange batch system SS on PP with N,R,a>0N,R,a>0.✦ The two normalizations used below are E*=Ra2E_*=Ra^2 and M=aNM=aN.

A dyadic time window J=(d,j)J=(d,j) is a pair with d≥0d\geq0 and 0≤j<2d0\leq j<2^d. Its time window interval is the cell [j/2d,(j+1)/2d)[j/2^d,(j+1)/2^d) of [0,1][0,1], with its right endpoint included when j=2d−1j=2^d-1. For a positive integer mm, the uniform partition of the unit interval has boundaries i/mi/m for 0≤i≤m0\leq i\leq m; its last cell is closed on the right and its other cells are open on the right. Batch bb has midpoint time tb=(b+1/2)/Rt_b=(b+1/2)/R.✦ For a finite list of times in an ordered set, the cut at xx is the number of times strictly below xx.✦ Thus the window starts at the cut at j/2dj/2^d, stops at the cut at (j+1)/2d(j+1)/2^d, and has batch count equal to the difference of these cuts. The time window system SJS_J is the consecutive batch system obtained by retaining those stages from the starting index through the stopping index, with the original slopes and marked sets.✦

Tokens and Rank Labels

Choose enumerations lb:Lb→ℤ/aℤl_b:L_b\to\mathbb Z/a\mathbb Z and gb:Gb→ℤ/aℤg_b:G_b\to\mathbb Z/a\mathbb Z. For a shift jj in the cyclic group ℤ/aℤ\mathbb Z/a\mathbb Z, the cyclic perfect matching is the bijection u↦gb−1(lb(u)+j)u\mapsto g_b^{-1}(l_b(u)+j). A matching between the differences of two finite sets induces a transport between the sets by fixing their common points and using the matching elsewhere; the inverse fixes the common points and uses the inverse matching.✦ For each shift jj, the batch token transport Ab→Ab+1A_b\to A_{b+1} is the bijection that fixes the common points and uses the cyclic perfect matching between the leaving and entering sets.✦ Cumulative token transport Tj,b:A0→AbT_{j,b}:A_0\to A_b starts with the identity and composes these transports through the first bb batches.✦

An edge token trace of e=(u,v)e=(u,v) consists of a batch bb, a matching number jj and an initial token t∈A0t\in A_0 with Tj,b(t)=uT_{j,b}(t)=u and Tj,b+1(t)=vT_{j,b+1}(t)=v, where u∈Lbu\in L_b and v∈Gbv\in G_b.✦ Choose one trace for each exchange edge.✦ Its batch determines the edge time t(e)=tbt(e)=t_b, and its matching number and initial token index determine the convex chain owner (j,t)(j,t).

A convex graph family of size mm consists of continuous convex functions gr:ℝ→ℝg_r:\mathbb R\to\mathbb R, 0≤r<m0\leq r<m, and a scale ε>0\varepsilon>0.✦ For a finite family of real graph functions grg_r, the vertical rank rank⁡g(x,y)\operatorname{rank}_g(x,y) counts the graphs strictly below (x,y)(x,y).✦ For mm such functions and a real scale ε\varepsilon, the smoothed rank coordinate qgq_g is the real-valued function in (2), where sigmoid\operatorname{sigmoid} is the sigmoid function; it is zero when m=0m=0. rank⁡g(x,y)=#{r:gr(x)<y},qg(x,y)=1m∑r=0m−1sigmoid⁡(y−gr(x)ε).\operatorname{rank}_g(x,y)=\#\{r:g_r(x)<y\}, \qquad q_g(x,y)=\frac1m\sum_{r=0}^{m-1} \operatorname{sigmoid}\!\left(\frac{y-g_r(x)}{\varepsilon}\right).(2) For m>0m>0, the unit interval rank coordinate is qgq_g, regarded as a point of [0,1][0,1].✦ The flat family is the convex graph family in which every graph equals zero and the scale is one.✦

A full convex chain graph extension of SS, relative to a finite planar set FF, consists of aNaN continuous convex graphs, one for each owner, and numbers ℓ,u\ell,u with s0<ℓ≤u<sRs_0<\ell\leq u<s_R. Their secant slopes lie in [ℓ,u][\ell,u], and every exchange segment lies on its owner’s graph. At a point of FF, a graph can meet that point only along an exchange segment owned by the graph. The window rank family is a chosen convex graph family of size aNaN for each SJS_J.✦ Write its unit interval coordinate as qJq_J. The rank label space ZZ and the configuration rank label zPz_P are defined in (3), with 𝒲\mathcal W denoting the set of dyadic windows. Z=[0,1]𝒲,zP(p)=(qJ(p))J∈𝒲.Z=[0,1]^{\mathcal W}, \qquad z_P(p)=(q_J(p))_{J\in\mathcal W}.(3)

Measures and Limits

A weighted atomic measure with weight w≥0w\geq0 on a finite labelled family (xi)(x_i) is w∑iδxiw\sum_i\delta_{x_i}.✦ For labels zvz_v in a measurable space XX, the normalized participating vertex measure αS,z\alpha_{S,z} and the weighted edge endpoint measure βS,z,E\beta_{S,z,E}, with nonnegative normalization EE, are given in (4). The weighted edge endpoint measure is zero when E=0E=0. αS,z=1N∑v∈V(S)δzv,βS,z,E=1E∑(u,v)∈ℰ(S)δzu.\alpha_{S,z}=\frac1N\sum_{v\in V(S)}\delta_{z_v}, \qquad \beta_{S,z,E}=\frac1E\sum_{(u,v)\in\mathcal E(S)}\delta_{z_u}.(4)

For a generic batch configuration, put zv=zP(Pv)z_v=z_P(P_v). The empirical point measure μP\mu^P, window vertex measure μJP\mu_J^P, edge label and time measure νP\nu^P, and its time restriction ρJP\rho_J^P are defined by (5). μP=1N∑v=0N−1δzv,μJP=αSJ,z,νP=1E*∑e=(u,v)∈ℰ(S)δ(zu,t(e)),ρJP=νP↾Z×J.\begin{aligned} \mu^P&=\frac1N\sum_{v=0}^{N-1}\delta_{z_v}, &\mu_J^P&=\alpha_{S_J,z},\\ \nu^P&=\frac1{E_*}\sum_{e=(u,v)\in\mathcal E(S)} \delta_{(z_u,t(e))}, &\rho_J^P&=\nu^P\!\restriction_{Z\times J}. \end{aligned}(5)

A straight-line drawing has a finite vertex set and a finite set of ordered edges, each directed from smaller to larger first coordinate and with both endpoints in its vertex set.✦ Two edges cross when their open segments intersect; crossing edge pairs count each unordered pair once. The exchange drawing uses V(S)V(S) and ℰ(S)\mathcal E(S).✦ For windows I,JI,J, the window crossing edge pairs 𝒞IJ\mathcal C_{IJ} are the crossing pairs with one exchange edge in each window. Choose a point pcp_c in the two open segments of each such pair.✦ The crossing label measure is σIJP=M−2∑c∈𝒞IJδzP(pc)\sigma_{IJ}^P=M^{-2}\sum_{c\in\mathcal C_{IJ}}\delta_{z_P(p_c)}.✦ For a straight-line drawing DD and a finite set EE of ordered point-index pairs, the exterior edge crossing points form the finite set of points where edges of DD cross outside the closed segments represented by EE.✦

Limiting rank data consist of finite measures μ,μJ\mu,\mu_J on ZZ, finite measures ν,ρJ\nu,\rho_J on Z×[0,1]Z\times[0,1], and finite crossing measures σIJ\sigma_{IJ} on ZZ.✦ Joint convergence of a sequence (Pr)(P_r) of configurations means weak convergence of every corresponding measure in (6), for every choice of windows. μPr⇒μ,μJPr⇒μJ,νPr⇒ν,ρJPr⇒ρJ,σIJPr⇒σIJ.\mu^{P_r}\Rightarrow\mu,\quad \mu_J^{P_r}\Rightarrow\mu_J,\quad \nu^{P_r}\Rightarrow\nu,\quad \rho_J^{P_r}\Rightarrow\rho_J,\quad \sigma_{IJ}^{P_r}\Rightarrow\sigma_{IJ}.(6)✦ The label marginals are λ=(πZ)*ν\lambda=(\pi_Z)_*\nu and λJ=(πZ)*ρJ\lambda_J=(\pi_Z)_*\rho_J, where πZ\pi_Z is projection onto ZZ. Write sJ=μJ(Z)s_J=\mu_J(Z) for the participating vertex mass.

Spatial Enhancement

The spatial cube is Q=[0,1]2Q=[0,1]^2.✦ The sigmoid compactification η\eta of the plane is the map η(x,y)=(sigmoid⁡(x),sigmoid⁡(y))\eta(x,y)=(\operatorname{sigmoid}(x),\operatorname{sigmoid}(y)) from ℝ2\mathbb R^2 to QQ.✦ Given any map ψ:ℝ2→ℝ2\psi:\mathbb R^2\to\mathbb R^2, the spatially enhanced rank label is the map zˆP:ℝ2→Z×Q\widehat z_P:\mathbb R^2\to Z\times Q with zˆP(p)=(zP(p),η(ψ(p)))\widehat z_P(p)=(z_P(p),\eta(\psi(p))).✦ The enhanced point measure and the enhanced window vertex and edge measures are defined by (7). μˆP=1N∑v=0N−1δzˆP(Pv),μˆJP=αSJ,zˆP∘P,λˆJP=βSJ,zˆP∘P,E*.\widehat\mu^P=\frac1N\sum_{v=0}^{N-1}\delta_{\widehat z_P(P_v)}, \qquad \widehat\mu_J^P=\alpha_{S_J,\widehat z_P\circ P}, \qquad \widehat\lambda_J^P=\beta_{S_J,\widehat z_P\circ P,E_*}.(7) The enhanced measure family consists of the point measure and the vertex and edge measures indexed by windows.✦ Convergence of the enhanced measure family means weak convergence of its point measure and every window vertex and edge measure, with limits denoted by μˆ,μˆJ,λˆJ\widehat\mu,\widehat\mu_J,\widehat\lambda_J.

The enhanced label space is the Cartesian product Z×QZ\times Q of the window label space and the spatial cube.✦

The Claim

For a sequence of generic batch configurations, write E*,r=Rrar2E_{*,r}=R_ra_r^2. The numerical hypotheses with constant C≥0C\geq0 are (8). 4608ar2Nr4E*,r3≤Cfor every r,13824Nr3E*,r3→0.\frac{4608a_r^2N_r^4}{E_{*,r}^3}\leq C\quad\text{for every }r, \qquad \frac{13824N_r^3}{E_{*,r}^3}\longrightarrow0.(8)

Theorem 1. Let a sequence of generic batch configurations satisfy (8) and converge jointly as in (6). For every dyadic window JJ and Borel set U⊂ZU\subset Z, the local cubic bound is (9). λJ(U)3≤CsJ3μJ(U).\lambda_J(U)^3\leq C\,s_J^3\mu_J(U).(9)✦ For each such window, the absolute continuity assertion is (10). λJ≪μ.\lambda_J\ll\mu.(10)✦

For any sequence of generic batch configurations satisfying (8), choose maps ψr\psi_r and assume convergence of the entire enhanced measure family. For every dyadic window JJ and Borel set U⊂Z×QU\subset Z\times Q, the enhanced local bound is (11). λˆJ(U)3≤CμˆJ(Z×Q)3μˆJ(U).\widehat\lambda_J(U)^3 \leq C\,\widehat\mu_J(Z\times Q)^3\widehat\mu_J(U).(11)✦

Let SS be a uniform exchange batch system on a generic NN-point configuration, with N,a>0N,a>0. Let XX be a measurable space with measurable singletons, assign labels zv∈Xz_v\in X, and fix E>0E>0. For every measurable set U⊂XU\subset X, the finite atomic bound is (12). βS,z,E(U)3≤4608a2N4E3(|V(S)|N)3αS,z(U)+13824|V(S)|3E3.\beta_{S,z,E}(U)^3 \leq \frac{4608a^2N^4}{E^3} \left(\frac{|V(S)|}{N}\right)^3\alpha_{S,z}(U) +\frac{13824|V(S)|^3}{E^3}.(12)✦

In the paper, the result reads: in the preceding limiting setup, every window JJ and Borel set U⊂ZU\subset Z satisfy λJ(U)3≤CsJ3μJ(U)\lambda_J(U)^3\leq C s_J^3\mu_J(U), where CC depends only on the fixed extremal comparison constants, in particular λ≪μ\lambda\ll\mu, and the estimate remains valid after labels in any compact metrizable space are attached to vertices and their edge endpoints and a further weak limit is taken.

What Is Checked

The claims are stated over generic planar batch systems with positive point count and exchange distance, joint weak limits satisfying (8), and convergent spatially enhanced measure families.