OctoInformBeta

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

Lemma 3.3: Absolute continuity for jointly convergent sweep batches with bounded crossing coefficient and vanishing errors

Crossing null sets control edge labels.

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

A rank sweep produces exchange edges when points leave and enter its marked set. The measures here record the rank labels of their initial endpoints and of the points where edges from different time windows cross.

An NN-point configuration is a map P:{0,…,N−1}→ℝ2P:\{0,\ldots,N-1\}\to\mathbb R^2. For planar points p,q,rp,q,r, the orientation determinant and the slope are defined by (1). orient⁡(p,q,r)=det⁡(q−p,r−p),s(p,q)=q2−p2q1−p1.\operatorname{orient}(p,q,r)=\det(q-p,r-p),\qquad s(p,q)=\frac{q_2-p_2}{q_1-p_1}.(1) A configuration PP is in general position when PP is injective and the orientation determinant is nonzero on every triple of distinct indices. A generic configuration PP is a configuration in general position whose first coordinates are pairwise distinct and whose unordered point pairs, represented by i<ji<j, have pairwise distinct slopes. A generic configuration PP also satisfies the following condition: a point in the common interior of two distinct point-pair segments belongs to no closed segment determined by a third point pair distinct from both.

The sweep projection πs\pi_s and the directed set distance dd between finite sets of point indices are given by (2). πs(p)=p2−sp1,d(A,B)=|A\B|.\pi_s(p)=p_2-sp_1,\qquad d(A,B)=|A\setminus B|.(2) A marked set AA is separated at slope ss when πs(Pi)<πs(Pj)\pi_s(P_i)<\pi_s(P_j) for every i∈Ai\in A and j∉Aj\notin A.✦ Uniform exchange batches of rank kk and distance aa are increasing sample slopes s0<⋯<sRs_0<\cdots<s_R with separated marked sets A0,…,ARA_0,\ldots,A_R, each of size kk, satisfying d(Ab,Ab+1)=ad(A_b,A_{b+1})=a.✦ A generic batch configuration consists of such a system on a generic configuration, with N,R,aN,R,a positive.✦

A rank kk sweep sequence has increasing sample slopes and separated marked sets of size kk, with directed distance at most one between successive marked sets.✦ The critical slopes of PP are the slopes of its distinct point pairs.✦ A critical slope traversal is an increasing list of cuts whose interior cuts are exactly those critical slopes; the canonical traversal lists them in order with one additional cut at each end. The intercritical sample slopes of a critical slope traversal are the midpoints of its consecutive cuts.✦ For generic PP and k≤Nk\le N, the canonical rank kk sweep is the rank sweep associated with this traversal.✦ A batch embedding into that sweep is a strictly increasing selection of its states with marked sets A0,…,ARA_0,\ldots,A_R; it requires k≤Nk\le N.✦

Write T=[0,1]T=[0,1]. A dyadic time window is a pair J=(d,q)J=(d,q), with d≥0d\ge0 and 0≤q<2d0\le q<2^d; let 𝒲\mathcal W be the set of these windows.✦ For a uniform partition into m>0m>0 cells, the boundary with index ii, from zero through mm, is i/mi/m, and each cell is left-closed and right-open except the last, which includes 11. The carrier |J|⊂T|J|\subset T of JJ is the cell in (3). |J|={[q/2d,(q+1)/2d),q+1<2d,[q/2d,1],q+1=2d.|J|= \begin{cases} [q/2^d,(q+1)/2^d),&q+1<2^d,\\ [q/2^d,1],&q+1=2^d. \end{cases}(3)✦ Batch bb has midpoint time τb\tau_b as in (4). For any finite family of ordered times tbt_b, the earlier-time count ct(x)c_t(x) is the number strictly below xx.✦ The starting index uJu_J, stopping index vJv_J, and batch count ℓJ\ell_J of a window are also defined in (4), where αJ=q/2d\alpha_J=q/2^d and βJ=(q+1)/2d\beta_J=(q+1)/2^d are its boundaries. τb=b+12R,ct(x)=#{b:tb<x},uJ=cτ(αJ),vJ=cτ(βJ),ℓJ=vJ−uJ.\begin{aligned} \tau_b&=\frac{b+\tfrac12}{R},& c_t(x)&=\#\{b:t_b<x\},\\ u_J&=c_\tau(\alpha_J),& v_J&=c_\tau(\beta_J),& \ell_J&=v_J-u_J. \end{aligned}(4) A consecutive batch window starting at uu with length ℓ\ell, where u+ℓ≤Ru+\ell\le R, retains the slopes and marked sets at stages uu through u+ℓu+\ell.✦ For a batch system SS, the time window system S|JS|J is the consecutive batch window with u=uJu=u_J and ℓ=ℓJ\ell=\ell_J.✦

Exchange Graphs and Rank Labels

The leaving and entering sets Lb,GbL_b,G_b, exchange edge set EE, and participating vertex set VV of a batch system SS are defined by (5). Lb=Ab\Ab+1,Gb=Ab+1\Ab,E=⋃b=0R−1(Lb×Gb),V={i:∃j,(i,j)∈E or (j,i)∈E}.\begin{aligned} L_b&=A_b\setminus A_{b+1},&G_b&=A_{b+1}\setminus A_b,\\ E&=\bigcup_{b=0}^{R-1}(L_b\times G_b),& V&=\{i:\exists j,\ (i,j)\in E\text{ or }(j,i)\in E\}. \end{aligned}(5) A straight-line graph drawing on PP is a finite vertex set and a finite set of ordered edges whose endpoints are vertices, each directed toward increasing first coordinate. The exchange graph drawing uses VV and EE.✦ Two edges cross when their open segments intersect.✦ The crossing edge pairs of a drawing count each unordered pair once, and a chosen crossing point lies in their common interior. Relative to a specified edge set, exterior crossing points are the drawing’s crossing points outside the closed segments of that set.✦

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 each cyclic shift jj, the batch perfect matching sends each i∈Lbi\in L_b to gb−1(lb(i)+j)g_b^{-1}(l_b(i)+j).✦ The matching-induced transport between marked sets fixes shared points and sends each leaving point to its matched entering point.✦ Cumulative token transport starts with the identity on A0A_0 and composes these transports through successive batches.✦ An edge token trace of e=(i,i′)e=(i,i') is a batch bb, a shift jj, and an initial token t∈A0t\in A_0 which occupies i∈Lbi\in L_b before the batch and i′∈Gbi'\in G_b after it.✦ A chosen trace assigns an edge its batch time τe=τb\tau_e=\tau_b and its convex chain owner (j,t)(j,t); the owner of a pair outside EE is (0,0)(0,0).

Convex chain graph extensions relative to a finite planar set HH are aNaN continuous convex functions grg_r on ℝ\mathbb R, with common secant slope bounds ℓ,u\ell,u satisfying s0<ℓ≤u<sRs_0<\ell\le u<s_R. Their secant bounds are those in (6), and every exchange edge lies on the graph assigned to its owner. A point of HH on a graph must lie on an exchange edge assigned to that graph. ℓ(y−x)≤gr(y)−gr(x)≤u(y−x)(x≤y).\ell(y-x)\le g_r(y)-g_r(x)\le u(y-x)\qquad(x\le y).(6)

A convex graph family of size mm is a family of continuous convex functions g0,…,gm−1:ℝ→ℝg_0,\ldots,g_{m-1}:\mathbb R\to\mathbb R with a positive scale ε\varepsilon.✦ The flat convex graph family (g,ε)(g,\varepsilon) is a family of zero functions with scale one.✦ For graph functions grg_r, the vertical rank counts the graphs strictly below a point, and its smoothed coordinate uses the real sigmoid function, as in (7). slot⁡g(p)=#{r:gr(p1)<p2},rg,ε(p)=1m∑r=0m−1sigmoid⁡(p2−gr(p1)ε).\operatorname{slot}_g(p)=\#\{r:g_r(p_1)<p_2\},\qquad r_{g,\varepsilon}(p)=\frac1m\sum_{r=0}^{m-1} \operatorname{sigmoid}\!\left(\frac{p_2-g_r(p_1)}{\varepsilon}\right).(7) The smoothed coordinate is defined for any real scale, with division by zero assigned value zero; it is zero for m=0m=0. For a convex graph family (g,ε)(g,\varepsilon) of positive size, the unit-interval rank coordinate rg,ε(p)r_{g,\varepsilon}(p) of a planar point pp is its smoothed coordinate and belongs to (0,1)(0,1).✦

The window rank graph family (gJ,εJ)(g_J,\varepsilon_J) of SS at J∈𝒲J\in\mathcal W is a convex family of size aNaN. The window label space is ℒ=T𝒲\mathcal L=T^{\mathcal W}.✦ The configuration rank label z:ℝ2→ℒz:\mathbb R^2\to\mathcal L assigns to each point its window rank coordinates, as in (8). z(p)(J)=rgJ,εJ(p).z(p)(J)=r_{g_J,\varepsilon_J}(p).(8)✦

Measures and Their Limits

For a finite family xhx_h in a measurable space and a nonnegative weight ww, the weighted atomic measure is w∑hδxhw\sum_h\delta_{x_h}.✦ The edge normalization is E*=Ra2E_*=Ra^2, and the chain normalization is M=aNM=aN. For windows I,JI,J, let 𝒞IJ\mathcal C_{IJ} be the exchange drawing’s crossing edge pairs with one edge in each window, and write cwc_w for the chosen crossing point of w∈𝒞IJw\in\mathcal C_{IJ}.

The empirical point label measure μF\mu^F and participating-vertex measure μJF\mu^F_J of a generic batch configuration FF use weight 1/N1/N per point. The edge label and time measure νF\nu^F uses initial-endpoint labels and batch times; its time restriction at JJ is νJF\nu^F_J. The crossing label measure σIJF\sigma^F_{IJ} applies zz to the window crossing points. These measures are given by (9). μF=1N∑i=0N−1δz(Pi),μJF=1N∑i∈V(S|J)δz(Pi),νF=1E*∑e∈Eδ(z(Pe1),τe),νJF=νF|ℒ×|J|,σIJF=1M2∑w∈𝒞IJδz(cw).\begin{aligned} \mu^F&=\frac1N\sum_{i=0}^{N-1}\delta_{z(P_i)},& \mu^F_J&=\frac1N\sum_{i\in V(S|J)}\delta_{z(P_i)},\\ \nu^F&=\frac1{E_*}\sum_{e\in E}\delta_{(z(P_{e_1}),\tau_e)},& \nu^F_J&=\nu^F\big|_{\mathcal L\times|J|},\\ \sigma^F_{IJ}&=\frac1{M^2}\sum_{w\in\mathcal C_{IJ}}\delta_{z(c_w)}. \end{aligned}(9) More generally, the window crossing point measure uses the last formula with any measurable-space labeling of planar points in place of zz; the normalized vertex measure and label and time measure likewise allow any labeling of point indices.

Limiting rank measures are finite measures μ,μJ,σIJ\mu,\mu_J,\sigma_{IJ} on ℒ\mathcal L and ν,νJ\nu,\nu_J on ℒ×T\mathcal L\times T, indexed by the indicated windows.✦ A sequence FmF_m of generic batch configurations converges jointly to these measures when all the weak convergences in (10) hold, for every fixed window or pair of windows. μFm⇒μ,μJFm⇒μJ,νFm⇒ν,νJFm⇒νJ,σIJFm⇒σIJ.\begin{aligned} \mu^{F_m}&\Rightarrow\mu,& \mu^{F_m}_J&\Rightarrow\mu_J,& \nu^{F_m}&\Rightarrow\nu,\\ \nu^{F_m}_J&\Rightarrow\nu_J,& \sigma^{F_m}_{IJ}&\Rightarrow\sigma_{IJ}. \end{aligned}(10)✦ The window edge label measure λJ\lambda_J is the label marginal of νJ\nu_J.✦ Writing pr⁡ℒ\operatorname{pr}_{\mathcal L} for the label projection, its definition is (11). λJ=(pr⁡ℒ)#νJ.\lambda_J=(\operatorname{pr}_{\mathcal L})_\#\nu_J.(11)✦ For a window KK, the disjoint subwindow pairs form the countable set 𝒫(K)\mathcal P(K) in (12). 𝒫(K)={(I,J)∈𝒲2:|I|⊆|K|,|J|⊆|K|,|I|∩|J|=⌀}.\mathcal P(K)=\{(I,J)\in\mathcal W^2: |I|\subseteq|K|,\ |J|\subseteq|K|,\ |I|\cap|J|=\varnothing\}.(12)✦

The Claim

Theorem 1. Let FmF_m be generic batch configurations with parameters Nm,Rm,km,amN_m,R_m,k_m,a_m, each admitting a batch embedding into its canonical rank sweep. Suppose their measures converge jointly as in (10) and Rm→∞R_m\to\infty. Put E*,m=Rmam2E_{*,m}=R_ma_m^2. Assume that for some C≥0C\ge0 the bounds in (13) hold, with the inequality valid for every mm. 4608am2Nm4E*,m3≤C,13824Nm3E*,m3→0,NmRmam→0.\frac{4608a_m^2N_m^4}{E_{*,m}^3}\le C,\qquad \frac{13824N_m^3}{E_{*,m}^3}\longrightarrow0,\qquad \frac{N_m}{R_ma_m}\longrightarrow0.(13) For every dyadic window KK, the absolute continuity statement (14) holds. λK≪∑(I,J)∈𝒫(K)σIJ.\lambda_K\ll\sum_{(I,J)\in\mathcal P(K)}\sigma_{IJ}.(14)✦

In the paper, the result reads: in the same setup, for every window KK, λK≪∑σIJ\lambda_K\ll\sum\sigma_{IJ} over disjoint dyadic subwindows I,JI,J of KK, where the sum specifies null sets, equivalently every Borel set BB null for all these crossing measures satisfies λK(B)=0\lambda_K(B)=0.

What Is Checked

The claim is stated over generic uniform exchange batches embedded in canonical rank sweeps, their joint measure limits, and the bounds in (13) with Rm→∞R_m\to\infty. For positive exchange distance, every exchange edge has a token trace.✦