OctoInformBeta

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

Lemma 4.3: For jointly convergent sweep batches satisfying the scale bounds, where the rank coordinates are differentiable

Transverse subwindows at almost every active label.

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 planar rank sweep marks the points with the smallest values of the projection y−sxy-sx as the slope ss increases. Grouping its exchanges into batches gives a straight-line graph and rank coordinates indexed by dyadic time windows. The question concerns the derivatives of these coordinates inside a window carrying positive conditional edge mass.

An NN-point configuration is a map P:{0,…,N−1}→ℝ2P:\{0,\ldots,N-1\}\to\mathbb R^2. A configuration PP in general position is an NN-point configuration in which PP is injective and no three distinct points are collinear, with orientation determinant det⁡(Pj−Pi,Pl−Pi)\det(P_j-P_i,P_l-P_i).✦ A generic configuration PP is an NN-point configuration in general position with 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, satisfying the following intersection condition: a point in the open segments of two distinct unordered pairs lies on no closed segment of a third distinct pair.✦

Batches and Time Windows

A marked set MM separated at slope ss is a set of point indices of a configuration PP for which Pi,2−sPi,1<Pj,2−sPj,1P_{i,2}-sP_{i,1}<P_{j,2}-sP_{j,1} for every i∈Mi\in M and j∉Mj\notin M.✦ A rank kk sweep sequence consists of increasing sample slopes and separated kk-element marked sets, with directed set distance d(M,M′)=|M\M′|d(M,M')=|M\setminus M'| at most one between successive sets.✦ The critical slopes are the slopes of distinct point pairs.✦ A traversal of the critical slopes of a configuration is an increasing list of cuts whose interior cuts are those slopes, with one cut below and one above them; intercritical samples are the midpoints of successive cuts. For a generic configuration and k≤Nk\le N, the canonical sweep is the rank kk sweep on this canonical traversal.✦

Uniform exchange batches on PP consist of increasing slopes s0<⋯<sRs_0<\cdots<s_R and separated marked sets M0,…,MRM_0,\ldots,M_R satisfying (1).✦ The leaving and entering sets of batch bb are LbL_b and GbG_b, respectively. |Mb|=k,Lb=Mb\Mb+1,Gb=Mb+1\Mb,|Lb|=a.|M_b|=k,\qquad L_b=M_b\setminus M_{b+1},\qquad G_b=M_{b+1}\setminus M_b,\qquad |L_b|=a.(1) A generic batch configuration is a generic point map on NN points with N,R,a>0N,R,a>0 and a system of uniform exchange batches satisfying (1).✦ An embedding into the canonical sweep is a strictly increasing selection of sweep states whose marked sets are M0,…,MRM_0,\ldots,M_R; it requires k≤Nk\le N.✦

The exchange edges form ℰ=⋃bLb×Gb\mathcal E=\bigcup_b L_b\times G_b, and the exchange vertices are all their endpoints. A straight-line drawing has a finite vertex set and ordered edges with endpoints in that set, each directed towards larger first coordinate; the exchange drawing uses these exchange vertices and edges. Two edges cross when their open segments intersect.✦ The crossing pairs of a straight-line drawing are its unordered pairs of crossing edges, each counted once, and a crossing point is a chosen point of their common interior. Exterior crossing points relative to an edge set are crossing points outside its closed segments.✦

Choose enumerations of each LbL_b and GbG_b by the cyclic indices of size aa. The cyclic perfect matching with index jj matches indices differing by jj modulo aa.✦ The transport induced by a cyclic perfect matching between successive marked sets is the bijection that fixes their common points and applies the matching to the leaving points.✦ For a fixed matching index jj and a stage bb, cumulative transport is the bijection from M0M_0 to MbM_b obtained by composing the batch transports with matching index jj through the first bb batches; at stage zero it is the identity.✦ An edge token trace records a batch, a matching index and a point of the initial marked set, called an initial token, whose positions immediately before and after the batch are the leaving and entering endpoints of the edge.✦ The convex chain owner of an exchange edge is the matching index and initial token index of a chosen edge token trace, with owner (0,0)(0,0) assigned to a nonedge.✦

Convex chain graph extensions, relative to a finite set QQ of planar points, are aNaN continuous convex functions gi:ℝ→ℝg_i:\mathbb R\to\mathbb R and bounds s0<ℓ≤u<sRs_0<\ell\le u<s_R satisfying (2). Every exchange segment lies on its owner’s graph, and every point of QQ on a graph belongs to an exchange segment with that owner. ℓ(y−x)≤gi(y)−gi(x)≤u(y−x)(x≤y).\ell(y-x)\le g_i(y)-g_i(x)\le u(y-x)\qquad(x\le y).(2)

Put I=[0,1]I=[0,1]. A dyadic window J=(d,j)J=(d,j) is a pair of integers with d≥0d\ge0 and 0≤j<2d0\le j<2^d; its time interval TJT_J is the uniform partition cell in (3). For a dyadic window KK, a subwindow of KK is a dyadic window whose time interval is contained in TKT_K.✦ T(d,j)={[j2−d,(j+1)2−d),j<2d−1,[j2−d,1],j=2d−1,𝒲(K)={J:TJ⊆TK}.T_{(d,j)}= \begin{cases} [j2^{-d},(j+1)2^{-d}),&j<2^d-1,\\ [j2^{-d},1],&j=2^d-1, \end{cases} \qquad \mathcal W(K)=\{J:T_J\subseteq T_K\}.(3) Batch bb has midpoint time tb=(b+12)/Rt_b=(b+\tfrac12)/R.✦ The starting and stopping indices of JJ count the batch times strictly below its left and right boundaries, respectively. The batch count of a dyadic window JJ is its stopping index minus its starting index.✦ A consecutive batch window with start uu and length vv, where u+v≤Ru+v\le R, retains the original slopes and marked sets at indices u,…,u+vu,\ldots,u+v; the time window system uses these starting and stopping indices. An edge’s time is the midpoint time of the batch in its chosen token trace.✦

Rank Labels and Limiting Measures

A convex graph family of size mm consists of continuous convex functions gi:ℝ→ℝg_i:\mathbb R\to\mathbb R and a scale ε>0\varepsilon>0.✦ The flat family is a convex graph family with gi=0g_i=0 and ε=1\varepsilon=1.✦ The vertical graph rank of a convex graph family at p=(p1,p2)p=(p_1,p_2) counts those graphs strictly below pp.✦ The smoothed rank coordinate and the vertical rank are given by (4); for an empty family the smoothed coordinate is zero. r(p)=#{i:gi(p1)<p2},f(p)=1m∑i=0m−1sigmoid⁡(p2−gi(p1)ε).r(p)=\#\{i:g_i(p_1)<p_2\},\qquad f(p)=\frac1m\sum_{i=0}^{m-1} \operatorname{sigmoid}\!\left(\frac{p_2-g_i(p_1)}{\varepsilon}\right).(4) The window rank graph family of a generic batch configuration has m=aNm=aN graphs. For a generic batch configuration, the rank label ℓ(p)\ell(p) is the function defined by ℓ(p)(J)=fJ(p)\ell(p)(J)=f_J(p) for every dyadic window JJ, where fJf_J is the unit interval rank coordinate of that window’s rank graph family.✦ The resulting rank label ℓ(p)\ell(p) belongs to the window label space 𝒵=I𝒲\mathcal Z=I^{\mathcal W}, where 𝒲\mathcal W is the set of all dyadic windows.

A weighted atomic measure is w∑q∈Aδxqw\sum_{q\in A}\delta_{x_q} for a finite index set AA, a nonnegative weight ww and points xqx_q in a measurable space.✦ For a batch configuration, put E*=Ra2E_*=Ra^2 and M=aNM=aN. Let VJV_J be the vertices participating in its time window system, and let 𝒞UV\mathcal C_{UV} be the crossing pairs with an edge in each of the windows U,VU,V, with chosen crossing point qwq_w. The empirical point label measure μfin\mu^{\mathrm{fin}} and participating-vertex measure μJfin\mu_J^{\mathrm{fin}}, together with the edge-time measure νfin\nu^{\mathrm{fin}}, its time restrictions ρJfin\rho_J^{\mathrm{fin}} and the crossing label measures σUVfin\sigma_{UV}^{\mathrm{fin}}, are defined by (5); the vertex normalization uses the full point count. μfin=1N∑v=0N−1δℓ(Pv),μJfin=1N∑v∈VJδℓ(Pv),νfin=1E*∑e∈ℰδ(ℓ(Pe1),te),ρJfin=νfin|𝒵×TJ,σUVfin=1M2∑w∈𝒞UVδℓ(qw).\begin{aligned} \mu^{\mathrm{fin}}&=\frac1N\sum_{v=0}^{N-1}\delta_{\ell(P_v)},& \mu_J^{\mathrm{fin}}&=\frac1N\sum_{v\in V_J}\delta_{\ell(P_v)},\\ \nu^{\mathrm{fin}}&=\frac1{E_*}\sum_{e\in\mathcal E} \delta_{(\ell(P_{e_1}),t_e)},& \rho_J^{\mathrm{fin}}&=\nu^{\mathrm{fin}}\big|_{\mathcal Z\times T_J},\\ \sigma_{UV}^{\mathrm{fin}}&=\frac1{M^2} \sum_{w\in\mathcal C_{UV}}\delta_{\ell(q_w)}. \end{aligned}(5)

Limiting rank data consist of finite measures μ,μJ,ν,ρJ\mu,\mu_J,\nu,\rho_J and σUV\sigma_{UV} on the respective spaces in (5).✦ A sequence of generic batch configurations has joint convergence to these data when all five weak convergences in (6) hold, for every choice of windows.✦ μ(m)⇒μ,μJ(m)⇒μJ,ν(m)⇒ν,ρJ(m)⇒ρJ,σUV(m)⇒σUV.\mu^{(m)}\Rightarrow\mu,\quad \mu_J^{(m)}\Rightarrow\mu_J,\quad \nu^{(m)}\Rightarrow\nu,\quad \rho_J^{(m)}\Rightarrow\rho_J,\quad \sigma_{UV}^{(m)}\Rightarrow\sigma_{UV}.(6)✦ Here superscript (m)(m) denotes the finite measure of the mmth configuration.

The edge label marginal is λ=(π𝒵)#ν\lambda=(\pi_{\mathcal Z})_\#\nu, where π𝒵(z,t)=z\pi_{\mathcal Z}(z,t)=z.✦ The conditional time law κz\kappa_z is the conditional kernel of ν\nu from 𝒵\mathcal Z to II; write hz(J)=κz(TJ)h_z(J)=\kappa_z(T_J).✦

Differentiable Coordinates

The signed coordinate pairs of (x,y)(x,y) are (x,−y)(x,-y) and (−x,y)(-x,y).✦ For fixed windows A,BA,B and a label z∈𝒵z\in\mathcal Z, use the first signed coordinate pair to define the rank coordinate point pzp_z in (7).✦ Given a window KK, let GJ:ℝ2→IG_J:\mathbb R^2\to I be a function for each J∈𝒲(K)J\in\mathcal W(K). At points of differentiability, DGJ(p)DG_J(p) denotes its real derivative, a linear form on ℝ2\mathbb R^2. pz=(z(A),−z(B))∈[0,1]×[−1,0].p_z=(z(A),-z(B))\in[0,1]\times[-1,0].(7)✦

For the mmth batch configuration, denote its parameters by Nm,Rm,amN_m,R_m,a_m, and put E*,m=Rmam2E_{*,m}=R_ma_m^2 and Mm=amNmM_m=a_mN_m. The scale bounds used in the rank assertion are (8), with a constant C0≥0C_0\ge0. Rm→∞,Mm→∞,4608am2Nm4E*,m3≤C0for every m,13824Nm3E*,m3→0,NmRmam→0.\begin{gathered} R_m\longrightarrow\infty,\qquad M_m\longrightarrow\infty,\qquad \frac{4608a_m^2N_m^4}{E_{*,m}^3}\le C_0\quad\text{for every }m,\\ \frac{13824N_m^3}{E_{*,m}^3}\longrightarrow0,\qquad \frac{N_m}{R_ma_m}\longrightarrow0. \end{gathered}(8)

The Claim

Theorem 1. Let ZZ be a measurable space with a measure η\eta, let w:Z→[0,∞]w:Z\to[0,\infty] be measurable, and let {τi:i∈𝒜}\{\tau_i:i\in\mathcal A\} be a countable family of measures. Let ci:Z→ℝ2c_i:Z\to\mathbb R^2 be measurable coordinate maps. Suppose the domination conditions (9) hold, where ℒ2\mathcal L^2 is planar Lebesgue measure. wη≪∑i∈𝒜τi,(ci)#τi≪ℒ2(i∈𝒜).w\eta\ll\sum_{i\in\mathcal A}\tau_i,\qquad (c_i)_\#\tau_i\ll\mathcal L^2\quad(i\in\mathcal A).(9) Let p:Z→ℝ2p:Z\to\mathbb R^2, let Hi:ℝ2→ℝ2H_i:\mathbb R^2\to\mathbb R^2, and let S⊆ZS\subseteq Z. Assume that, for every z∈Sz\in S and i∈𝒜i\in\mathcal A, ci(z)=Hi(p(z))c_i(z)=H_i(p(z)) and HiH_i is differentiable at p(z)p(z). For η\eta-almost every z∈Sz\in S with w(z)≠0w(z)\ne0, ∃i∈𝒜:det⁡DHi(p(z))≠0.\exists i\in\mathcal A:\qquad \det DH_i(p(z))\ne0.(10)✦

Let a sequence of generic batch configurations have embeddings into their canonical sweeps, joint convergence (6), and the bounds (8). Fix dyadic windows K,A,BK,A,B and functions GJG_J as above. For λ\lambda-almost every label zz at which every GJG_J, J∈𝒲(K)J\in\mathcal W(K), is differentiable at pzp_z and satisfies GJ(pz)=z(J)G_J(p_z)=z(J), each J∈𝒲(K)J\in\mathcal W(K) with hz(J)≠0h_z(J)\ne0 satisfies ∃C,D∈𝒲(J),TC∩TD=⌀,[v↦(DGC(pz)v,DGD(pz)v)]:ℝ2→ℝ2is injective.\begin{gathered} \exists C,D\in\mathcal W(J),\qquad T_C\cap T_D=\varnothing,\\ \left[v\longmapsto \bigl(DG_C(p_z)v,\,DG_D(p_z)v\bigr)\right] :\mathbb R^2\longrightarrow\mathbb R^2 \quad\text{is injective}. \end{gathered}(11)✦

In the paper, the result reads: in the preceding positive chart on a set EE of labels, after deleting a further λ|E\lambda|_E-null set, every z∈Ez\in E has the property that each dyadic J⊆KJ\subseteq K with hz(J)>0h_z(J)>0 contains disjoint dyadic subwindows C,DC,D with det⁡(∇GC(pz),∇GD(pz))≠0\det(\nabla G_C(p_z),\nabla G_D(p_z))\ne0.

What Is Checked

The nonsingularity assertion (10) concerns countably many measurable planar coordinates with the domination conditions (9); the rank assertion (11) concerns the joint limits of generic sweep batch configurations and differentiable subwindow coordinates at the signed rank point (7).

Every exchange edge has a token trace.✦ The unit interval rank coordinate of a convex graph family of positive size lies strictly between zero and one.