OctoInformBeta

OpenAI preprint, September 2026 · §4. Differentiable coordinates and transverse directions

§4 (setting): the positive chart and allowed directions: Subsequence limits for countable window families, with admissible projective parameters at simultaneous differentiability points in the reflected chart

Differentiable rank limits in signed coordinates.

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

Two time windows supply a pair of rank coordinates for a planar sweep. The question is whether the other window ranks have differentiable limits in these coordinates, and which directions their differentials permit at a given time.

An indexed configuration of NN points is a map P:{0,…,N−1}→ℝ2P:\{0,\ldots,N-1\}\to\mathbb R^2. An indexed configuration PP of points in ℝ2\mathbb R^2 is in general position when its points are distinct and the orientation determinant det⁡(Pj−Pi,Pl−Pi)\det(P_j-P_i,P_l-P_i) is nonzero for three distinct indices.✦ A generic configuration PP is an indexed configuration in general position that has pairwise distinct first 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. It also satisfies the following intersection condition: for three distinct unordered point pairs, a point in the open segments of the first two pairs never belongs to the closed segment of the third. Analytic notions on ℝ2\mathbb R^2 use the maximum norm on the product.

A system SS of RR uniform exchange batches at rank kk and distance aa consists of sample slopes sbs_b and marked index sets ℳb\mathcal M_b, for 0≤b≤R0\le b\le R, satisfying (1).✦ Projection separation compares the sweep projection πs(x,y)=y−sx\pi_s(x,y)=y-sx of every marked point with every unmarked point. s0<⋯<sR,|ℳb|=k,πsb(Pi)<πsb(Pj)(i∈ℳb,j∉ℳb),ℒb:=ℳb\ℳb+1,ℋb:=ℳb+1\ℳb,|ℒb|=a(0≤b<R).\begin{gathered} s_0<\cdots<s_R,\qquad |\mathcal M_b|=k,\\ \pi_{s_b}(P_i)<\pi_{s_b}(P_j) \quad(i\in\mathcal M_b,\ j\notin\mathcal M_b),\\ \mathcal L_b:=\mathcal M_b\setminus\mathcal M_{b+1},\qquad \mathcal H_b:=\mathcal M_{b+1}\setminus\mathcal M_b,\qquad |\mathcal L_b|=a\quad(0\le b<R). \end{gathered}(1) The sets ℒb\mathcal L_b and ℋb\mathcal H_b in (1) are the leaving and entering points. The exchange graph has edge set E(S)E(S) and participating vertex set V(S)V(S) defined by (2). E(S)=⋃b=0R−1ℒb×ℋb,V(S)={i:∃j,(i,j)∈E(S)or(j,i)∈E(S)}.E(S)=\bigcup_{b=0}^{R-1}\mathcal L_b\times\mathcal H_b, \qquad V(S)=\{i:\exists j,\ (i,j)\in E(S)\ \text{or}\ (j,i)\in E(S)\}.(2) A straight-line graph drawing on PP is a finite vertex set and a finite set of ordered edges whose endpoints lie in that vertex set, each edge oriented toward larger first coordinate. The exchange drawing of SS on PP is the straight-line graph drawing with edge set E(S)E(S) and vertex set V(S)V(S) given in (2). Two edges cross when their open segments intersect; crossing pairs count each unordered pair once. A crossing point qwq_w of a pair ww of crossing edges is a chosen point in the intersection of their open segments.✦ For a specified edge set E0E_0, the exterior crossing points are the drawing’s crossing points outside the closed segments of E0E_0.✦

For N,a>0N,a>0, enumerate each leaving and entering set by 0,…,a−10,\ldots,a-1. The cyclic perfect matching of shift jj matches a leaving index rr to the entering index r+jr+j modulo aa. The batch token transport for a cyclic perfect matching in a batch of SS is the map that fixes shared marked points and uses the matching on leaving points; the cumulative token transport from the initial marked set of SS is the composition of the batch token transport maps. An edge token trace for an exchange edge ee is a choice of a batch, a shift and an initial token occupying the source just before that batch and the target just after it, with the source leaving and the target entering.✦ Fix a chosen trace for each exchange edge. The convex chain owner of an exchange edge is the pair consisting of the shift and initial token index of its chosen trace, with owner (0,0)(0,0) for a pair outside E(S)E(S).✦ The leaving point enumeration of each batch is a chosen bijection from its leaving points to the indices 0,…,a−10,\ldots,a-1. The entering point enumeration of each batch is a chosen bijection from its entering points to the indices 0,…,a−10,\ldots,a-1.

For N,a>0N,a>0 and a finite test set H⊂ℝ2H\subset\mathbb R^2, a convex chain graph extension (gr)(g_r) of SS is a family of aNaN continuous convex functions on ℝ\mathbb R, indexed by chain owners, with bounds ℓ≤u\ell\le u satisfying s0<ℓs_0<\ell and u<sRu<s_R. Every secant satisfies ℓ(y−x)≤gr(y)−gr(x)≤u(y−x)\ell(y-x)\le g_r(y)-g_r(x)\le u(y-x) for x≤yx\le y. Each exchange segment lies on its owner’s graph, and any point of HH on a graph lies on an exchange segment with that owner.

Window Ranks and Their Measures

Put I=[0,1]I=[0,1]. A dyadic time window J=(d,j)J=(d,j) is a pair of indices with d≥0d\ge0 and 0≤j<2d0\le j<2^d; its endpoints and carrier are given in (3).✦ These are cells of the uniform partition, whose boundaries for mm cells are i/mi/m, 0≤i≤m0\le i\le m. For dyadic time windows AA and BB, write A≺BA\prec B when uA≤ℓBu_A\le\ell_B.✦ ℓJ=j2d,uJ=j+12d,J={[ℓJ,uJ],j=2d−1,[ℓJ,uJ),j<2d−1.\ell_J=\frac{j}{2^d},\qquad u_J=\frac{j+1}{2^d},\qquad J=\begin{cases} [\ell_J,u_J],&j=2^d-1,\\ [\ell_J,u_J),&j<2^d-1. \end{cases}(3) For R>0R>0, the batch midpoint times are τb=(b+1/2)/R\tau_b=(b+1/2)/R.✦ The starting and stopping indices of a dyadic time window JJ in a batch system are the numbers of batch midpoint times strictly below ℓJ\ell_J and uJu_J, respectively. The batch count of a dyadic time window JJ is the difference between its stopping and starting indices, truncated at zero.✦ The time window system S|JS|J for a batch system SS and a dyadic time window JJ is the consecutive batch window of SS starting at the starting index of JJ and having the batch count of JJ as its length; its samples are the corresponding consecutive samples of SS.✦ Denote its edges and vertices by EJE_J and VJV_J. The time τe\tau_e of an exchange edge is the midpoint time of its chosen trace’s batch.✦ For dyadic time windows II and JJ, let WIJW_{IJ} be the crossing edge pairs with one edge in EIE_I and one in EJE_J, and let qwq_w be their crossing points in the exchange drawing.

A convex graph family of size mm consists of mm continuous convex functions gr:ℝ→ℝg_r:\mathbb R\to\mathbb R and a scale ε>0\varepsilon>0.✦ For a convex graph family (gr)(g_r) with scale ε>0\varepsilon>0, the vertical graph rank rank⁡g\operatorname{rank}_g and smoothed rank coordinate zg,εz_{g,\varepsilon} are functions on ℝ2\mathbb R^2 given in (4), where sigmoid\operatorname{sigmoid} denotes the logistic sigmoid. For m>0m>0 the smoothed coordinate lies in (0,1)(0,1) and is the family’s II-valued rank coordinate; for m=0m=0 its value is zero. The flat family of size mm is the convex graph family whose graphs are all zero and whose scale is one.✦ rank⁡g(p)=#{r:gr(p1)<p2},zg,ε(p)=1m∑r=0m−1sigmoid⁡(p2−gr(p1)ε).\operatorname{rank}_g(p)=\#\{r:g_r(p_1)<p_2\},\qquad z_{g,\varepsilon}(p)=\frac1m\sum_{r=0}^{m-1} \operatorname{sigmoid}\!\left(\frac{p_2-g_r(p_1)}{\varepsilon}\right).(4) A generic batch configuration FF consists of a generic PP and a batch system with N,R,a>0N,R,a>0.✦ For a generic batch configuration FF and a dyadic time window JJ, the window rank graph family of FF at JJ is a convex graph family of size aNaN. The configuration rank label zF(p)z_F(p) of a point pp for a generic batch configuration FF is the function assigning to each dyadic time window JJ the rank coordinate of the window rank graph family of FF at JJ, evaluated at pp.✦ Thus zF(p)z_F(p) belongs to the window label space 𝒵=I𝒲\mathcal Z=I^{\mathcal W}, where 𝒲\mathcal W is the set of dyadic windows; write zJz_J for the JJ coordinate of a label zz.✦

For a finite family (xr)(x_r) in a measurable space, the weighted atomic measure is w∑rδxrw\sum_r\delta_{x_r}, with w≥0w\ge0.✦ A vertex labeling gives the normalized vertex measure by weighting each participating vertex by 1/N1/N; its label and time measure weights each edge by E*−1E_*^{-1} at its source label and edge time, where E*=Ra2E_*=Ra^2. For dyadic windows I,JI,J and a map z:ℝ2→Xz:\mathbb R^2\to X to a measurable space, the window crossing point measure is the finite measure on XX with weight M−2M^{-2} at z(qw)z(q_w) for each crossing pair w∈WIJw\in W_{IJ}, where M=aNM=aN. For rank labels, the empirical point measure μF\mu_F, window vertex measures μF,J\mu_{F,J}, edge-time measure νF\nu_F, time restrictions νF,J\nu_{F,J} and crossing label measures σF,IJ\sigma_{F,IJ} are the measures in (5). μF=1N∑v=0N−1δzF(Pv),μF,J=1N∑v∈VJδzF(Pv),νF=1E*∑e∈E(S)δ(zF(Pe1),τe),νF,J=νF|𝒵×J,σF,IJ=1M2∑w∈WIJδzF(qw).\begin{aligned} \mu_F&=\frac1N\sum_{v=0}^{N-1}\delta_{z_F(P_v)},& \mu_{F,J}&=\frac1N\sum_{v\in V_J}\delta_{z_F(P_v)},\\ \nu_F&=\frac1{E_*}\sum_{e\in E(S)} \delta_{(z_F(P_{e_1}),\tau_e)},& \nu_{F,J}&=\nu_F|_{\mathcal Z\times J},\\ \sigma_{F,IJ}&=\frac1{M^2}\sum_{w\in W_{IJ}}\delta_{z_F(q_w)}. \end{aligned}(5) A limiting rank measure family consists of finite measures μ,μJ,σIJ\mu,\mu_J,\sigma_{IJ} on 𝒵\mathcal Z and ν,νJ\nu,\nu_J on 𝒵×I\mathcal Z\times I.✦ A sequence (Fm)(F_m) of generic batch configurations converges jointly to a limiting rank measure family when all the weak convergences in (6) hold, for every window JJ and pair I,JI,J.✦ The label marginal of ν\nu is denoted by λ\lambda. μFm⇒μ,μFm,J⇒μJ,νFm⇒ν,νFm,J⇒νJ,σFm,IJ⇒σIJ.\mu_{F_m}\Rightarrow\mu,\qquad \mu_{F_m,J}\Rightarrow\mu_J,\qquad \nu_{F_m}\Rightarrow\nu,\qquad \nu_{F_m,J}\Rightarrow\nu_J,\qquad \sigma_{F_m,IJ}\Rightarrow\sigma_{IJ}.(6)

Signed Coordinates and Direction Parameters

For windows A,BA,B, the signed coordinate pairs p+(z)p^+(z) and p−(z)p^-(z) are the points in (7).✦ The sign ++ will correspond to windows before both AA and BB, and −- to windows after both. Write QQ for the first coordinate reflection and Ω\Omega for the reflected rank chart in (7). p+(z)=(zA,−zB),p−(z)=(−zA,zB),Q(x,y)=(−x,y),Ω=(0,1)×(−1,0).p^+(z)=(z_A,-z_B),\qquad p^-(z)=(-z_A,z_B),\qquad Q(x,y)=(-x,y),\qquad \Omega=(0,1)\times(-1,0).(7) For maps ρ:X→ℝ2\rho:X\to\mathbb R^2 and c:X→Ic:X\to I, their monotone rank extension Eρ,c:ℝ2→IE_{\rho,c}:\mathbb R^2\to I is the supremum in (8), with the coordinatewise order.✦ For a generic batch configuration FmF_m, a dyadic time window CC and a sign ϵ\epsilon, the configuration rank extension Em,CϵE^\epsilon_{m,C} is the monotone rank extension formed from the signed rank pair of FmF_m and its CC coordinate.✦ Eρ,c(x)=sup⁡({0}∪{c(r):r∈X,ρ(r)≤x}),Em,Cϵ=Epϵ∘zFm,r↦zFm(r)C,ϵ∈{+,−}.\begin{aligned} E_{\rho,c}(x)&=\sup\bigl(\{0\}\cup\{c(r):r\in X,\ \rho(r)\le x\}\bigr),\\ E^\epsilon_{m,C}&= E_{p^\epsilon\circ z_{F_m},\ r\mapsto z_{F_m}(r)_C}, \qquad \epsilon\in\{+,-\}. \end{aligned}(8)

Definition 1. Let TT be an invertible linear map of ℝ2\mathbb R^2, let (Li)(L_i) be real linear forms, and let αi,βi\alpha_i,\beta_i be real time endpoints. The admissible projective parameter set 𝒟T(L,α,β;t)\mathcal D_T(L,\alpha,\beta;t) for the linear map TT, linear forms (Li)(L_i) and time endpoints (αi),(βi)(\alpha_i),(\beta_i) at time tt is the subset of [0,1][0,1] defined by (9), using the simplex representative s(x)=(x,1−x)s(x)=(x,1-x). 𝒟T(L,α,β;t)={x∈[0,1]:∀i,βi<t⇒Li(Ts(x))≥0,t<αi⇒Li(Ts(x))≤0}.\mathcal D_T(L,\alpha,\beta;t)= \left\{x\in[0,1]:\ \forall i,\quad \begin{aligned} \beta_i<t&\ \Longrightarrow\ L_i(Ts(x))\ge0,\\ t<\alpha_i&\ \Longrightarrow\ L_i(Ts(x))\le0 \end{aligned} \right\}.(9)✦

The Claim

Theorem 1. Let (Fm)(F_m) be generic batch configurations with batch counts RmR_m, converging jointly to a limiting rank measure family, with Rm→∞R_m\to\infty. Let A≺BA\prec B and let (Ci)(C_i) be a countable family of windows. Choose ϵ=+\epsilon=+ with Ci≺A,BC_i\prec A,B for every ii, or ϵ=−\epsilon=- with A,B≺CiA,B\prec C_i for every ii. There are strictly increasing indices ϕ:ℕ→ℕ\phi:\mathbb N\to\mathbb N and functions Gi:ℝ2→IG_i:\mathbb R^2\to I satisfying (10) for every ii. Giis coordinatewise monotone and measurable,Giis differentiable at Lebesgue almost every p,zCi=Gi(pϵ(z))(z∈supp⁡μ,Gi continuous at pϵ(z)),limm→∞x→pEϕ(m),Ciϵ(x)=Gi(p)(Gi continuous at p).\begin{gathered} G_i\ \text{is coordinatewise monotone and measurable},\\ G_i\ \text{is differentiable at Lebesgue almost every }p,\\ z_{C_i}=G_i(p^\epsilon(z)) \quad\bigl(z\in\operatorname{supp}\mu,\quad G_i\text{ continuous at }p^\epsilon(z)\bigr),\\ \lim_{\substack{m\to\infty\\x\to p}} E^\epsilon_{\phi(m),C_i}(x)=G_i(p) \quad\bigl(G_i\text{ continuous at }p\bigr). \end{gathered}(10)✦

Let (Fm)(F_m) instead be any sequence of generic batch configurations with Rm→∞R_m\to\infty, and let A≺BA\prec B and (Ci)(C_i) be any family with Ci≺A,BC_i\prec A,B. Suppose Gi:ℝ2→IG_i:\mathbb R^2\to I is differentiable almost everywhere for each ii and satisfies the joint limit in (10) along the full sequence with ϵ=+\epsilon=+. At a point p∈Ωp\in\Omega where every GiG_i is differentiable, and for t∈It\in I with t≤ℓA,ℓBt\le\ell_A,\ell_B, the conclusion is (11), where dGi(p)dG_i(p) is the differential at pp. 𝒟Q((dGi(p))i,(ℓCi)i,(uCi)i;t)≠⌀.\mathcal D_Q\bigl((dG_i(p))_i,(\ell_{C_i})_i,(u_{C_i})_i;t\bigr) \ne\varnothing.(11)✦

In the paper, the result reads: choose separated dyadic windows K,K′⊂(0,1)K,K'\subset(0,1) and a set EE of positive λ\lambda-measure where hz(K),hz(K′)>0h_z(K),h_z(K')>0 for the conditional time measures hzh_z, then restrict further so that λ|E≪σAB\lambda|_E\ll\sigma_{AB} for disjoint A,B⊂K′A,B\subset K'; in the chart p=(zA,zB)p=(z_A,z_B), the bounded coordinatewise monotone limits fCf_C, for C⊂KC\subset K, are differentiable and satisfy zC=fC(p)z_C=f_C(p) after null sets are discarded; for tt in the interior of KK, the set Dp(t)D_p(t) consists of unit directions dd in the common signed quadrant satisfying ∇fC(p)⋅d≥0\nabla f_C(p)\cdot d\ge0 for CC strictly before tt and ∇fC(p)⋅d≤0\nabla f_C(p)\cdot d\le0 for CC strictly after tt, with strict separation requiring a positive time gap.

What Is Checked

The rank limit assertion is stated over joint weak limits of generic batch configurations and their signed rank extensions; the direction assertion is stated over the full sequence of extensions and simultaneous differentiability points in the reflected chart. The exchange edges satisfy the endpoint condition of the exchange drawing✦, and, for a>0a>0, every exchange edge has a token trace.✦