OctoInformBeta

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

Lemma 3.2: Joint convergence of the window measures, with sweep embeddings and explicit bounds on the edge normalization

Atomless time laws from normalized sweep limits.

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

For a chosen rank kk, an increasing slope sweep at slope ss selects the lowest kk points in the projection y−sxy-sx of a point with planar coordinates (x,y)(x,y). Sampling the sweep in batches records exchanges of marked points. Each exchange edge receives a time and a label consisting of its rank coordinates in all dyadic time windows. The limiting edge measure has a conditional time law at each label.

The plane is ℝ2\mathbb R^2, and a configuration of NN points is an indexed map P:{0,…,N−1}→ℝ2P:\{0,\ldots,N-1\}\to\mathbb R^2. For points p,q,rp,q,r in ℝ2\mathbb R^2, the orientation determinant and the slope of the pair p,qp,q are given 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 in ℝ2\mathbb R^2 is in general position if its points are distinct and no three are collinear.✦ A generic configuration PP in ℝ2\mathbb R^2 is a configuration in general position that has distinct horizontal coordinates and distinct slopes for distinct unordered pairs, and satisfies the following condition: for three distinct unordered pairs, represented by i<ji<j, a<ba<b and c<dc<d, a common point of the first two open segments does not belong to the third closed segment.✦

A system of uniform exchange batches on PP, with parameters R,k,aR,k,a, consists of strictly increasing slopes s0,…,sRs_0,\ldots,s_R and marked index sets B0,…,BRB_0,\ldots,B_R satisfying (2).✦ Projection separation means the strict inequality there✦; the directed set distance from AA to BB is d(A,B)=|A\B|d(A,B)=|A\setminus B|.✦ A generic batch configuration has N,R,a>0N,R,a>0, a generic point configuration and such a batch system.✦ |Bb|=k,(Pi)2−sb(Pi)1<(Pj)2−sb(Pj)1(i∈Bb,j∉Bb),d(Bb,Bb+1)=a.|B_b|=k,\qquad (P_i)_2-s_b(P_i)_1<(P_j)_2-s_b(P_j)_1 \quad(i\in B_b,\ j\notin B_b),\qquad d(B_b,B_{b+1})=a.(2)

The critical slopes are the slopes of all distinct point pairs.✦ A critical slope traversal of a finite set CC of real slopes is a strictly increasing list of cuts whose interior cuts are precisely the members of CC.✦ The canonical critical slope traversal of a finite set CC of real slopes lists CC in increasing order between one additional cut below all its members and one above them.✦ In any critical slope traversal, each intercritical sample slope is the midpoint of two consecutive cuts.✦ A rank kk sweep sequence of a configuration PP is a sequence of strictly increasing sample slopes together with separating marked sets of size kk, with directed set distance at most one between successive marked sets.✦ For a generic configuration and k≤Nk\le N, the canonical rank sweep has one more state than the number of interior cuts. Writing BjsweepB^{\mathrm{sweep}}_j for the marked set at sweep state jj, a batch embedding ι\iota of a rank kk batch system into its canonical rank sweep is a strictly increasing map from the batch sample indices to sweep states, with the mark matching in (3); the embedding also requires k≤Nk\le N.✦ Bb=Bι(b)sweep(0≤b≤R).B_b=B^{\mathrm{sweep}}_{\iota(b)}\qquad(0\le b\le R).(3)

Time Windows and Exchange Edges

Write I=[0,1]I=[0,1]. A dyadic time window J=(d,j)J=(d,j) is a pair of integers with depth d≥0d\ge0 and 0≤j<2d0\le j<2^d.✦ For m>0m>0, the uniform partition boundaries are i/mi/m✦; its cells and the carrier IJI_J of a window are defined by (4). Cm,i={[i/m,(i+1)/m),i<m−1,[i/m,1],i=m−1,IJ=C2d,j.C_{m,i}= \begin{cases} [i/m,(i+1)/m),&i<m-1,\\ [i/m,1],&i=m-1, \end{cases} \qquad I_J=C_{2^d,j}.(4) The batch midpoint time is tb=(b+12)/Rt_b=(b+\tfrac12)/R.✦ For any finite ordered family of times tt, let ct(x)=|{b:tb<x}|c_t(x)=|\{b:t_b<x\}| be its number of earlier times.✦ The window starting index qJq_J, stopping index rJr_J and batch count ℓJ\ell_J are given by (5), with subtraction truncated at zero. A consecutive batch window of a batch system, starting at qq with length ℓ\ell and satisfying q+ℓ≤Rq+\ell\le R, is the subsystem consisting of the slopes and marked sets at indices q,…,q+ℓq,\ldots,q+\ell✦; the time window system uses q=qJq=q_J and ℓ=ℓJ\ell=\ell_J.✦ qJ=ct(j/2d),rJ=ct((j+1)/2d),ℓJ=rJ−qJ.q_J=c_t(j/2^d),\qquad r_J=c_t((j+1)/2^d),\qquad \ell_J=r_J-q_J.(5)

The leaving and entering index sets of batch bb are LbL_b and GbG_b in (6). The exchange edges form the set EE there, and the participating vertices form VV. Put E*=Ra2E_*=Ra^2 and M=aNM=aN for the edge and chain normalizations, respectively. Lb=Bb\Bb+1,Gb=Bb+1\Bb,E=⋃b<R(Lb×Gb),V={i:∃j,(i,j)∈E or (j,i)∈E}.L_b=B_b\setminus B_{b+1},\qquad G_b=B_{b+1}\setminus B_b,\qquad E=\bigcup_{b<R}(L_b\times G_b),\qquad V=\{i:\exists j,\ (i,j)\in E\text{ or }(j,i)\in E\}.(6)

Chosen enumerations of LbL_b and GbG_b by the cyclic indices modulo aa give a perfect matching for each shift uu: an index numbered vv is matched to the entering index numbered v+uv+u.✦ The transport induced by a matching between two marked sets is the map that fixes their shared points and sends each leaving point to its matched entering point.✦ For a fixed shift uu and a stage 0≤b≤R0\le b\le R, the cumulative token transport from B0B_0 to BbB_b is the bijection obtained by composing the first bb batch transports with shift uu, starting with the identity at stage zero.✦ An edge token trace for an exchange edge e=(i,j)e=(i,j) is a choice of a batch bb, a shift uu and an initial token whose positions just before and after that batch are ii and jj, with i∈Lbi\in L_b and j∈Gbj\in G_b.✦ Choose one trace for every exchange edge.✦ For an exchange edge ee, the batch in its chosen trace determines its edge time tet_e, and the shift and initial token in that trace determine its convex chain owner. The owner of a pair outside EE is (0,0)(0,0).

A straight-line graph drawing is a finite graph drawn by straight segments, with each edge oriented towards increasing horizontal coordinate and both endpoints in the vertex set.✦ The exchange drawing uses VV and EE.✦ Two edges cross when their open segments intersect.✦ Crossing edge pairs count each unordered pair once✦, and a chosen crossing point lies in their common interior.✦ The set 𝒞AB\mathcal C_{AB} of crossing pairs for windows AA and BB consists of unordered pairs of crossing exchange edges with one edge in each window✦; write xwx_w for the chosen crossing point of a pair w∈𝒞ABw\in\mathcal C_{AB}.✦ The exterior crossing points of a straight-line graph drawing relative to an edge set are its crossing points outside the closed segments represented by that set.✦

Window Rank Labels

For a generic batch configuration with chain normalization MM and a finite test set H⊂ℝ2H\subset\mathbb R^2, full convex chain graph extensions consist of a family of MM continuous convex functions gi:ℝ→ℝg_i:\mathbb R\to\mathbb R and bounds ℓ,u\ell,u satisfying (7). Each exchange segment lies on the graph assigned by its convex chain owner. Every point of HH on a graph belongs to an exchange segment assigned to that graph. s0<ℓ≤u<sR,ℓ(y−x)≤gi(y)−gi(x)≤u(y−x)(x≤y).s_0<\ell\le u<s_R,\qquad \ell(y-x)\le g_i(y)-g_i(x)\le u(y-x)\quad(x\le y).(7)

A convex graph family of size mm consists of continuous convex functions g0,…,gm−1g_0,\ldots,g_{m-1} on ℝ\mathbb R and a scale ε>0\varepsilon>0.✦ The flat convex graph family of size mm consists of mm zero functions and has scale one.✦ For arbitrary real graph functions gig_i and a real scale ε\varepsilon, the vertical graph rank rg(p)r_g(p) counts graphs strictly below pp✦, and the smoothed rank coordinate ug,ε(p)u_{g,\varepsilon}(p) uses the real sigmoid, as in (8).✦ The smoothed coordinate is zero when m=0m=0. For a convex graph family (g,ε)(g,\varepsilon) of positive size, the unit-interval rank coordinate of a point pp is the smoothed coordinate ug,ε(p)u_{g,\varepsilon}(p) regarded as a point of II.✦ rg(p)=|{i:gi(p1)<p2}|,ug,ε(p)=1m∑i=0m−1sigmoid⁡(p2−gi(p1)ε).r_g(p)=|\{i:g_i(p_1)<p_2\}|,\qquad u_{g,\varepsilon}(p)=\frac1m\sum_{i=0}^{m-1} \operatorname{sigmoid}\!\left(\frac{p_2-g_i(p_1)}{\varepsilon}\right).(8)

Each dyadic window JJ of a generic batch configuration has an associated convex rank graph family (gJ,εJ)(g^J,\varepsilon_J) of size MM. The window label space Z=I𝒲Z=I^{\mathcal W} consists of functions from the set 𝒲\mathcal W of dyadic windows to II.✦ The configuration rank label zP:ℝ2→Zz_P:\mathbb R^2\to Z is the window rank label defined by (9).✦ zP(p)(J)=ugJ,εJ(p).z_P(p)(J)=u_{g^J,\varepsilon_J}(p).(9)

Measures and Limits

A weighted atomic measure η\eta on a measurable space XX is a measure of the form w∑iδxiw\sum_i\delta_{x_i}, where the points xix_i form a finite indexed family in XX and ww is a nonnegative weight.✦ Write VJV_J for the participating vertices of the time window system. The empirical point label measure μP\mu^P, window vertex label measure μJP\mu_J^P, edge label and time measure νP\nu^P, time-restricted edge measure νJP\nu_J^P, and crossing label measure σABP\sigma_{AB}^P are the finite measures in (10). The normalized vertex measure and the label and time measure allow arbitrary vertex labels in a measurable space; the former is zero when N=0N=0. The window crossing point measure allows an arbitrary planar labeling in the last sum. μP=1N∑i=0N−1δzP(Pi),μJP=1N∑i∈VJδzP(Pi),νP=1E*∑e=(i,j)∈Eδ(zP(Pi),te),νJP=νP|Z×IJ,σABP=1M2∑w∈𝒞ABδzP(xw).\begin{aligned} \mu^P&=\frac1N\sum_{i=0}^{N-1}\delta_{z_P(P_i)},& \mu_J^P&=\frac1N\sum_{i\in V_J}\delta_{z_P(P_i)},\\ \nu^P&=\frac1{E_*}\sum_{e=(i,j)\in E}\delta_{(z_P(P_i),t_e)},& \nu_J^P&=\nu^P\big|_{Z\times I_J},\\ \sigma_{AB}^P&=\frac1{M^2}\sum_{w\in\mathcal C_{AB}}\delta_{z_P(x_w)}. \end{aligned}(10)

Limiting rank data ℒ\mathcal L consist of finite measures μ,μJ\mu,\mu_J and σAB\sigma_{AB} on ZZ, and ν,νJ\nu,\nu_J on Z×IZ\times I.✦ A sequence FmF_m of generic batch configurations converges jointly to ℒ\mathcal L when all five families converge weakly as specified in (11), for every fixed choice of windows.✦ μFm⇒μ,μJFm⇒μJ,νFm⇒ν,νJFm⇒νJ,σABFm⇒σAB.\mu^{F_m}\Rightarrow\mu,\qquad \mu_J^{F_m}\Rightarrow\mu_J,\qquad \nu^{F_m}\Rightarrow\nu,\qquad \nu_J^{F_m}\Rightarrow\nu_J,\qquad \sigma_{AB}^{F_m}\Rightarrow\sigma_{AB}.(11) The edge label marginal λ\lambda is the pushforward in (12).✦ The conditional time kernel z↦hzz\mapsto h_z of ν\nu is the family of conditional time laws on II obtained by conditioning ν\nu on its label in ZZ.✦ λ=(πZ)#ν,πZ(z,t)=z.\lambda=(\pi_Z)_\#\nu,\qquad \pi_Z(z,t)=z.(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, admitting batch embeddings into their canonical rank sweeps and converging jointly to limiting rank data ℒ\mathcal L. Suppose Rm→∞R_m\to\infty. Set E*,m=Rmam2E_{*,m}=R_ma_m^2, and suppose there is a constant C≥0C\ge0 such that 4608am2Nm4/E*,m3≤C4608a_m^2N_m^4/E_{*,m}^3\le C for every mm, while 13824Nm3/E*,m3→013824N_m^3/E_{*,m}^3\to0. Then hz({t})=0for every t∈I,for λ-almost every z∈Z.h_z(\{t\})=0\quad\text{for every }t\in I, \qquad\text{for }\lambda\text{-almost every }z\in Z.(13)✦

In the paper, the result reads: in the same extremal limiting setup, hzh_z is nonatomic for λ\lambda-almost every label zz.

What Is Checked

The claim (13) is stated over generic uniform exchange configurations, their sweep embeddings and the joint limits of their window measures, under the specified coefficient bound and vanishing error. The unit-interval rank coordinate of a positive convex graph family lies strictly between zero and one.