OctoInformBeta

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

§3 (setting): rank coordinates and the extremal limiting setup: Dyadic windows index the graph families, while separated slope bounds determine a paired coordinate chart

Rank coordinates from convex graph families.

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 slope sweep follows the lowest projected points through successive exchanges. Convex graph families give a continuous version of vertical rank, and dyadic time windows index the resulting labels.

A planar point is a pair of real numbers.✦ An NN-point configuration is an indexed map P:{0,…,N−1}→ℝ2P:\{0,\ldots,N-1\}\to\mathbb R^2. A configuration PP is in general position if its points are distinct and det⁡(Pj−Pi,Pℓ−Pi)≠0\det(P_j-P_i,P_\ell-P_i)\ne0 for distinct indices i,j,ℓi,j,\ell; this determinant is their orientation. A generic configuration PP is a configuration in general position with distinct horizontal coordinates and distinct slopes (Pj,2−Pi,2)/(Pj,1−Pi,1)(P_{j,2}-P_{i,2})/(P_{j,1}-P_{i,1}) for distinct unordered pairs i<ji<j, such that no intersection of the open segments of two distinct unordered pairs lies on the closed segment of a third distinct unordered pair.✦

The sweep projection at slope ss is qs(x,y)=y−sxq_s(x,y)=y-sx.✦ A slope ss separates a marked index set KK when qs(Pi)<qs(Pj)q_s(P_i)<q_s(P_j) for i∈Ki\in K and j∉Kj\notin K.✦ A uniform exchange system SS consists of sample slopes and marked sets satisfying (1); it has RR batches, rank kk and exchange distance aa. s0<⋯<sR,|Kb|=k,sb separates Kb(0≤b≤R),|Kb\Kb+1|=a(0≤b<R).s_0<\cdots<s_R,\qquad |K_b|=k,\qquad s_b\text{ separates }K_b\quad(0\le b\le R),\qquad |K_b\setminus K_{b+1}|=a\quad(0\le b<R).(1)✦ The leaving and entering sets of batch bb are Lb=Kb\Kb+1L_b=K_b\setminus K_{b+1} and Gb=Kb+1\KbG_b=K_{b+1}\setminus K_b. The exchange graph has edge set E=⋃b=0R−1Lb×GbE=\bigcup_{b=0}^{R-1}L_b\times G_b✦ and vertex set VV consisting of their endpoints.✦ A consecutive batch window of SS is the system obtained by restricting its stages to c,…,c+ℓc,\ldots,c+\ell, where c+ℓ≤Rc+\ell\le R.✦

Exchange Graphs and Convex Families

For a batch with leaving set LL, entering set GG and exchange distance a>0a>0, enumerations ℓ:L→ℤ/aℤ\ell:L\to\mathbb Z/a\mathbb Z and γ:G→ℤ/aℤ\gamma:G\to\mathbb Z/a\mathbb Z define the cyclic perfect matching u↦γ−1(ℓ(u)+j)u\mapsto\gamma^{-1}(\ell(u)+j) for each j∈ℤ/aℤj\in\mathbb Z/a\mathbb Z.✦ The leaving and entering point enumerations apply this construction to each batch, giving its jjth perfect matching. A matching between the differences of two finite marked sets induces a bijection between those sets by fixing their shared points and matching the others.✦ The batch token transport Tbj:Kb→Kb+1T_b^j:K_b\to K_{b+1} is the bijection that fixes the shared points of the two marked sets and matches their remaining points by the jjth batch matching.✦ The cumulative token transport Ubj:K0→KbU_b^j:K_0\to K_b is the bijection obtained by starting with the identity and successively composing the batch token transports.✦

An edge token trace (b,j,t)(b,j,t) for e=(u,v)e=(u,v) consists of a batch, a matching index and an initial token t∈K0t\in K_0 with Ubj(t)=u∈LbU_b^j(t)=u\in L_b and Ub+1j(t)=v∈GbU_{b+1}^j(t)=v\in G_b.✦ The convex chain owner of an exchange edge is the pair (j,t)(j,t) from its chosen edge token trace; its initial token is regarded as a point index.✦

A straight-line drawing of PP is a finite vertex set and a finite set of ordered edges, each directed toward the larger horizontal coordinate and with both endpoints in the vertex set.✦ The exchange graph drawing uses VV and EE.✦ Two edges have an interior intersection when their open endpoint segments intersect.✦ The crossing edge pairs of a drawing are the unordered pairs of edges whose open endpoint segments intersect✦, and a chosen crossing point is a point in the corresponding intersection.✦ For a drawing DD and a finite edge set E0E_0 on the same configuration, the exterior edge crossing points relative to E0E_0 form the finite set of chosen crossing points of DD lying on none of the closed segments specified by E0E_0.✦

A convex graph family 𝒢∈𝔉m\mathcal G\in\mathfrak F_m consists of continuous convex functions gi:ℝ→ℝg_i:\mathbb R\to\mathbb R, indexed by 0≤i<m0\le i<m, and a positive scale ε\varepsilon.✦ The flat family in 𝔉m\mathfrak F_m is the convex graph family with gi=0g_i=0 and scale 11.✦ For N,a>0N,a>0 and a finite set F⊂ℝ2F\subset\mathbb R^2, convex chain graph extensions of SS are aNaN continuous convex graphs with secant slopes between constants s0<l≤u<sRs_0<l\le u<s_R. Every exchange segment lies on the graph indexed by its convex chain owner. A point of FF on one of the graphs must lie on an exchange segment owned by that graph.

Slope Bounds and Vertical Levels

The vertical graph rank of a point (x,y)(x,y) in a finite graph family is the number #{i:gi(x)<y}\#\{i:g_i(x)<y\}.✦ The open unit interval is U=(0,1)U=(0,1).✦ A slope-bounded rank coordinate r∈𝒞(l,u)r\in\mathcal C(l,u) is a continuous function r:ℝ2→Ur:\mathbb R^2\to U whose restriction to each vertical line is strictly increasing and onto UU, and which satisfies (2). r(x+h,y+lh)≤r(x,y)≤r(x+h,y+uh)(x,y∈ℝ,h≥0).r(x+h,y+lh)\le r(x,y)\le r(x+h,y+uh) \qquad(x,y\in\mathbb R,\ h\ge0).(2)✦ The rank level height Lr(x,v)L_r(x,v), for v∈Uv\in U, is the unique height yy with r(x,y)=vr(x,y)=v.✦ For A∈𝒞(lA,uA)A\in\mathcal C(l_A,u_A) and B∈𝒞(lB,uB)B\in\mathcal C(l_B,u_B), the rank level height difference Δv\Delta_v at v=(vA,vB)∈U2v=(v_A,v_B)\in U^2 is the function given by Δv(x)=LA(x,vA)−LB(x,vB)\Delta_v(x)=L_A(x,v_A)-L_B(x,v_B).✦

The Definition

Definition 1 (Window families and rank coordinates). The dyadic time windows form the index set 𝒲\mathcal W in (3). 𝒲={(d,j):d∈ℕ,0≤j<2d}.\mathcal W=\{(d,j):d\in\mathbb N,\ 0\le j<2^d\}.(3)✦ The uniform partition of the closed unit interval into m>0m>0 parts has boundaries i/mi/m for 0≤i≤m0\le i\le m.✦ For R>0R>0, the batch midpoint times and the stages selected by J=(d,j)∈𝒲J=(d,j)\in\mathcal W are given in (4); the count ct(x)c_t(x) is the number of times strictly earlier than xx.✦ tb=b+1/2R(0≤b<R),ct(x)=#{b:tb<x},αJ=j2d,βJ=j+12d,pJ=ct(αJ),qJ=ct(βJ),ℓJ=qJ−pJ.\begin{gathered} t_b=\frac{b+1/2}{R}\quad(0\le b<R),\qquad c_t(x)=\#\{b:t_b<x\},\\ \alpha_J=\frac{j}{2^d},\qquad \beta_J=\frac{j+1}{2^d},\qquad p_J=c_t(\alpha_J),\qquad q_J=c_t(\beta_J),\qquad \ell_J=q_J-p_J. \end{gathered}(4) The consecutive system for JJ is the consecutive batch window of SS that starts at stage pJp_J and has ℓJ\ell_J batches. The window starting batch index pJp_J is the number of batch midpoint times strictly before the left boundary of JJ.✦ The window stopping batch index qJq_J is the number of batch midpoint times strictly before the right boundary of JJ.✦

Let Λ(t)=(1+e−t)−1\Lambda(t)=(1+e^{-t})^{-1} be the logistic function. For real functions gi:ℝ→ℝg_i:\mathbb R\to\mathbb R, 0≤i<m0\le i<m, and a real scale ε\varepsilon, the smoothed graph rank rg,εr_{g,\varepsilon} is the function in (5). rg,ε(x,y)={0,m=0,12,m>0,ε=0,1m∑i=0m−1Λ(y−gi(x)ε),m>0,ε≠0.r_{g,\varepsilon}(x,y)= \begin{cases} 0,&m=0,\\ \dfrac12,&m>0,\ \varepsilon=0,\\ \displaystyle\frac1m\sum_{i=0}^{m-1} \Lambda\!\left(\frac{y-g_i(x)}{\varepsilon}\right), &m>0,\ \varepsilon\ne0. \end{cases}(5)✦ Suppose m>0m>0, ε>0\varepsilon>0 and the gig_i are continuous, with the secant slope bounds in (6). l(x′−x)≤gi(x′)−gi(x)≤u(x′−x)(0≤i<m,x≤x′).l(x'-x)\le g_i(x')-g_i(x)\le u(x'-x) \qquad(0\le i<m,\ x\le x').(6) The smoothed rank coordinate Cg,εl,uC_{g,\varepsilon}^{l,u} is the member of 𝒞(l,u)\mathcal C(l,u) given in (7). Cg,εl,u∈𝒞(l,u),Cg,εl,u(x,y)=rg,ε(x,y).C_{g,\varepsilon}^{l,u}\in\mathcal C(l,u),\qquad C_{g,\varepsilon}^{l,u}(x,y)=r_{g,\varepsilon}(x,y).(7)✦ When slope-bounded rank coordinates A∈𝒞(lA,uA)A\in\mathcal C(l_A,u_A) and B∈𝒞(lB,uB)B\in\mathcal C(l_B,u_B) satisfy uA<lBu_A<l_B, their paired rank homeomorphism HA,BH_{A,B} is given by (8). HA,B:ℝ2→≅U2,HA,B(p)=(A(p),B(p)).H_{A,B}:\mathbb R^2\xrightarrow{\ \cong\ }U^2,\qquad H_{A,B}(p)=(A(p),B(p)).(8)✦ For a uniform exchange system SS on a generic configuration PP, with N,a,R>0N,a,R>0, and J∈𝒲J\in\mathcal W, the window convex graph family 𝒢S,J\mathcal G_{S,J} is the family in (9). 𝒢S,J=((gJ,i)0≤i<M,εJ)∈𝔉M,M=aN.\mathcal G_{S,J}= \bigl((g_{J,i})_{0\le i<M},\varepsilon_J\bigr)\in\mathfrak F_M, \qquad M=aN.(9)✦ The compact window label space ZZ is the countable cube in (10), with its product topology. Z=[0,1]𝒲={z:𝒲→[0,1]}.Z=[0,1]^{\mathcal W}=\{z:\mathcal W\to[0,1]\}.(10)✦

In the paper, the result reads: for a sequence of generic systems of RR consecutive distance-aa batches on NN points with N/a→∞N/a\to\infty and R≍(N/a)4/3R\asymp(N/a)^{4/3}, put E*=Ra2E_*=Ra^2 and M=aNM=aN, give the batches the times (i−1/2)/R(i-1/2)/R in dyadic windows, and for MM full convex graphs with slopes inside the span of JJ set rJ(x,y)=M−1∑i=1MΦ((y−gJ,i(x))/εJ)r_J(x,y)=M^{-1}\sum_{i=1}^M\Phi((y-g_{J,i}(x))/\varepsilon_J), where Φ\Phi is the distribution function of a continuous strictly positive probability density and εJ>0\varepsilon_J>0 is sufficiently small; for disjoint ordered windows A,BA,B, the map (rA,rB):ℝ2→(0,1)2(r_A,r_B):\mathbb R^2\to(0,1)^2 is a homeomorphism, and the labels z=(rJ)Jz=(r_J)_J lie in the countable cube ZZ.

What Is Checked

The window families are stated over generic finite configurations and uniform exchange systems with N,a,R>0N,a,R>0; the paired chart is stated over arbitrary slope-bounded rank coordinates with uA<lBu_A<l_B. For continuous graphs, the smoothed rank is continuous at every real scale.✦ For m>0m>0 its values belong to UU at every real scale.✦ For any nonempty finite family of real functions and any positive scale, the smoothed graph rank is strictly increasing with height on each vertical line.✦ For any nonempty finite family of real functions and any positive scale, the smoothed graph rank attains every level in (0,1)(0,1) on each vertical line.✦

For a>0a>0, every exchange edge has an edge token trace.✦

For separated slope bounds uA<lBu_A<l_B, each height difference Δv\Delta_v has a zero.✦ With uA<lBu_A<l_B and v∈U2v\in U^2, the rank level intersection abscissa X(v)X(v) is the chosen zero of Δv\Delta_v✦, and the inverse of the paired chart is given by (11). HA,B−1(v)=(X(v),LA(X(v),vA)).H_{A,B}^{-1}(v)=\bigl(X(v),L_A(X(v),v_A)\bigr).(11)