OctoInformBeta

OpenAI preprint, September 2026 · §7. Boundary activity and the packing contradiction

Lemma 7.2: Moving strips and finite budgets constrain spatially enhanced rank measures

Boundary control for enhanced edge 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

Fix a dyadic interval of a planar rank sweep and retain the points participating in its exchanges. The spatial question concerns their mass outside three parallel strips when both the rank labels and the spatial coordinates have weak limits.

An indexed configuration is a map P:{0,…,N−1}→ℝ2P:\{0,\ldots,N-1\}\to\mathbb R^2. An indexed configuration PP is in general position when it is injective and no three of its points are collinear; the orientation of three configuration points p,q,rp,q,r is det⁡(q−p,r−p)\det(q-p,r-p). An indexed configuration PP in the plane is generic when it is in general position, has distinct first coordinates and distinct slopes (q2−p2)/(q1−p1)(q_2-p_2)/(q_1-p_1) for distinct unordered pairs, and satisfies the following further condition: the common interior point of two segments from distinct unordered pairs belongs to no segment from a third distinct pair.✦ Unordered pairs are represented by i<ji<j.

Uniform exchange batches on PP consist of increasing slopes s0<⋯<sRs_0<\cdots<s_R and sets A0,…,ARA_0,\ldots,A_R of kk marked indices. At slope sbs_b, every marked point has smaller projection y−sbxy-s_bx than every unmarked point, and |Ab\Ab+1|=a|A_b\setminus A_{b+1}|=a. The leaving and entering sets of batch bb are Lb=Ab\Ab+1L_b=A_b\setminus A_{b+1} and Gb=Ab+1\AbG_b=A_{b+1}\setminus A_b. The exchange edges are ℰ=⋃bLb×Gb\mathcal E=\bigcup_b L_b\times G_b, and the participating vertices are their endpoints. A generic batch configuration FF has N,R,a>0N,R,a>0, a generic point map, and such a batch system.✦

A rank kk sweep of an indexed planar configuration PP is an increasing sequence of sample slopes with separated kk-element marked sets, whose successive directed set distances |A\A′||A\setminus A'| are at most one.✦ The critical slopes are all slopes of pairs of distinct configuration points.✦ Their canonical traversal lists them increasingly between one lower and one upper cut; each intercritical sample is the midpoint of consecutive cuts. The canonical rank sweep uses these samples. A batch embedding of a uniform exchange batch system on PP into the canonical rank kk sweep of PP is an increasing selection of sweep states with the same marked sets as the batch system and k≤Nk\le N.✦

A dyadic window J=(d,j)J=(d,j) is a pair consisting of a dyadic depth dd and a cell index jj with 0≤j<2d0\le j<2^d.✦ The interval IJI_J of a dyadic window J=(d,j)J=(d,j) is [j2−d,(j+1)2−d)[j2^{-d},(j+1)2^{-d}), with the last cell closed at 11.✦ Batch bb has time (b+12)/R(b+\tfrac12)/R.✦ The starting and stopping indices of JJ count batch times strictly below its left and right boundaries, respectively. The window batch system for a dyadic window JJ in a uniform exchange batch system is the consecutive sequence of batches between the starting and stopping indices of JJ; its length is the stopping index minus the starting index.

Graphs, Labels and Measures

Enumerations of LbL_b and GbG_b give aa cyclic perfect matchings, by adding a fixed shift modulo aa. A matching induces a transport between successive marked sets that fixes their intersection. Composing these transports carries an initial marked token through the batches.✦ A token trace for an exchange edge given by an ordered pair of point indices is a choice of batch, matching number and initial marked token; the token occupies the source before the batch and the target afterwards, and the source leaves the marked set while the target enters it.✦ The convex chain owner of an edge is its matching number and initial token index in a chosen trace.✦ For a fixed matching number, the cumulative token transport to stage bb is the bijection from the initial marked set to the marked set at bb obtained by composing the batch transports, starting with the identity at stage zero.

A full convex graph extension of a uniform exchange batch system is a family of aNaN continuous convex graphs containing the edges assigned to their chains. Its secant slopes lie between bounds ℓ≤u\ell\le u strictly inside the first and last batch slopes. Relative to a finite test set, every test point on a graph belongs to an edge assigned to that graph. The exterior crossing points of a collection of edges are the drawing’s crossing points outside those edges’ closed segments.✦

A convex graph family is a finite collection g1,…,gmg_1,\ldots,g_m of continuous convex functions on ℝ\mathbb R, together with a scale ε>0\varepsilon>0.✦ The sigmoid used in its smoothed rank coordinate is the logistic function. For a convex graph family g1,…,gmg_1,\ldots,g_m with scale ε\varepsilon, the vertical rank at a point pp is the number of graphs with gi(p1)<p2g_i(p_1)<p_2, and the smoothed rank coordinate rg(p)r_g(p) is given by (1). rg(p)=1m∑i=1msigmoid⁡(p2−gi(p1)ε).r_g(p)=\frac1m\sum_{i=1}^m \operatorname{sigmoid}\!\left(\frac{p_2-g_i(p_1)}{\varepsilon}\right).(1) The coordinate in (1) is zero for m=0m=0, and belongs to (0,1)(0,1) for m>0m>0. The flat family is a convex graph family whose functions are all zero and whose scale is one.✦ Each batch window has a rank graph family of M=aNM=aN graphs. The configuration label ℓF(p)\ell_F(p) of a point pp in a batch configuration FF is the element of the label space ℒ=[0,1]{dyadic windows}\mathcal L=[0,1]^{\{\text{dyadic windows}\}} whose JJ coordinate is the rank coordinate of the graph family for window JJ.

A weighted atomic measure is w∑iδxiw\sum_i\delta_{x_i} on a finite indexed family, with w≥0w\ge0.✦ Write 𝒱J\mathcal V_J and ℰJ\mathcal E_J for the vertices and edges of the window batch system. The finite label measures are defined by μF=N−1∑iδℓF(Pi),μF,J=N−1∑i∈𝒱JδℓF(Pi),νF=E*−1∑e∈ℰδ(ℓF(Pe1),te),E*=Ra2,M=aN.\begin{aligned} \mu_F&=N^{-1}\sum_i\delta_{\ell_F(P_i)},& \mu_{F,J}&=N^{-1}\sum_{i\in\mathcal V_J}\delta_{\ell_F(P_i)},\\ \nu_F&=E_*^{-1}\sum_{e\in\mathcal E} \delta_{(\ell_F(P_{e_1}),\,t_e)},& E_*&=Ra^2,\qquad M=aN. \end{aligned}(2) Here tet_e is the midpoint time of the edge’s batch.✦ The edge-time measure in (2) has time restriction νF,J=νF|ℒ×IJ\nu_{F,J}=\nu_F|_{\mathcal L\times I_J}.✦ Two edges cross when their open segments intersect; crossing pairs are counted once. For windows I,JI,J, the crossing label measure σF,IJ\sigma_{F,IJ} puts weight M−2M^{-2} at the label of the chosen interior crossing point of every crossing pair with one edge in each window.✦ For a labeling zz of the point indices into a measurable space XX and a nonnegative normalization EE, the weighted edge endpoint measure is the finite measure on XX with weight E−1E^{-1} at z(e1)z(e_1) for each exchange edge ee; it is zero when E=0E=0. For a labeling zz of an NN-point configuration into a measurable space XX, the normalized vertex measure is the finite measure on XX with weight 1/N1/N at the label of each participating vertex; it is zero when N=0N=0.

Joint convergence of configurations FnF_n to limiting rank measures means the weak convergences μFn→μ,μFn,J→μJ,νFn→ν,νFn,J→νJ,σFn,IJ→σIJfor every I,J.\mu_{F_n}\to\mu,\quad \mu_{F_n,J}\to\mu_J,\quad \nu_{F_n}\to\nu,\quad \nu_{F_n,J}\to\nu_J,\quad \sigma_{F_n,IJ}\to\sigma_{IJ} \quad\text{for every }I,J.(3)✦ In (3), the vertex and crossing measures live on ℒ\mathcal L, and the edge-time measures live on ℒ×[0,1]\mathcal L\times[0,1]. With πℒ\pi_{\mathcal L} denoting projection onto the label coordinate, the label marginal is λ=(πℒ)*ν\lambda=(\pi_{\mathcal L})_*\nu; the window edge label measure is λJ=(πℒ)*νJ\lambda_J=(\pi_{\mathcal L})_*\nu_J.

Spatial Coordinates and Strips

Put Q=[0,1]2Q=[0,1]^2.✦ The spatial compactification enc⁡:ℝ2→Q\operatorname{enc}:\mathbb R^2\to Q is the map that applies the sigmoid to both coordinates. The decoding map is dec⁡(v)=(−log⁡(v1−1−1),−log⁡(v2−1−1))\operatorname{dec}(v)=(-\log(v_1^{-1}-1),-\log(v_2^{-1}-1)). For an invertible linear map EE, an origin oo and a scale r>0r>0, the physical coordinate map is ψ(p)=rE−1(p−o)\psi(p)=rE^{-1}(p-o).✦ The enhanced label of pp is zF(p)=(ℓF(p),enc⁡(ψ(p)))∈ℒ×Qz_F(p)=(\ell_F(p),\operatorname{enc}(\psi(p)))\in\mathcal L\times Q.✦ The enhanced label space is the Cartesian product ℒ×Q\mathcal L\times Q of the window label space and the unit square.

The enhanced point measure μˆF\widehat\mu_F of a batch configuration FF is the atomic measure that puts weight N−1N^{-1} at every zF(Pi)z_F(P_i).✦ The enhanced participating-vertex measure μˆF,J\widehat\mu_{F,J} of a batch configuration FF in a window JJ is the atomic measure that puts weight N−1N^{-1} at zF(Pi)z_F(P_i) for every i∈𝒱Ji\in\mathcal V_J.✦ The enhanced edge measure λˆF,J\widehat\lambda_{F,J} of a batch configuration FF in a window JJ is the atomic measure that puts weight E*−1E_*^{-1} at zF(Pe1)z_F(P_{e_1}) for every e∈ℰJe\in\mathcal E_J.✦ The enhanced measure family of a batch configuration FF is a family consisting of one enhanced point measure and an enhanced participating-vertex measure and enhanced edge measure for each window.✦ The original label family has the corresponding entries μ,μJ,λJ\mu,\mu_J,\lambda_J.✦

For a metric label space XX, let q:X×Q→Qq:X\times Q\to Q be spatial projection and put ρ=dec⁡∘q\rho=\operatorname{dec}\circ q. For real parameters tt and θ\theta and a triple cc of real strip offsets, use the usual scalar product and normal n(t)=(−t,1−t)n(t)=(-t,1-t) to define the region outside three strips and its compactified lift by A(t,θ,c)={x∈(−2,2)2:|n(t)⋅x−ci|>18θ for i=1,2,3},UX(t,θ,c)=X×enc⁡(A(t,θ,c)).\begin{aligned} A(t,\theta,c)&=\{x\in(-2,2)^2: |n(t)\cdot x-c_i|>18\theta\text{ for }i=1,2,3\},\\ U_X(t,\theta,c)&=X\times\operatorname{enc}(A(t,\theta,c)). \end{aligned}(4) Norms and balls in the plane use the maximum norm, so closed balls are squares. Compact strip parameters form the space 𝒫=[0,4]×[0,1]×[0,1]×[−4,4]3\mathcal P=[0,4]\times[0,1]\times[0,1]\times[-4,4]^3, with coordinates (b,t,θ,c)(b,t,\theta,c).✦

The Claim

Theorem 1. Let XX be a metric space with its Borel sets.

Suppose finite measures αn\alpha_n on X×QX\times Q converge weakly to α\alpha, and tn→tt_n\to t, θn→θ\theta_n\to\theta, bn→bb_n\to b, cn→cc_n\to c in their real coordinate spaces. If eventually αn(UX(tn,θn,cn))≤2bn\alpha_n(U_X(t_n,\theta_n,c_n))\le2b_n, the moving-strip conclusion is (5): α(UX(t,θ,c))≤2b.\alpha(U_X(t,\theta,c))\le2b.(5)✦

Let α,η\alpha,\eta be finite measures on X×QX\times Q with η≤α\eta\le\alpha, and let HH be measurable. Suppose ∥ρ(z)∥≤1\|\rho(z)\|\le1 and enc⁡(ρ(z))=q(z)\operatorname{enc}(\rho(z))=q(z) on HH. Take C,b,θ≥0C,b,\theta\ge0, d>1d>1, t∈ℝt\in\mathbb R and c∈ℝ3c\in\mathbb R^3. Assume ρ*(α|H)(B¯(x,s))≤Csd\rho_*(\alpha|_H)(\overline B(x,s))\le Cs^d for every x∈ℝ2x\in\mathbb R^2 and s>0s>0, and η(UX(t,θ,c))≤2b\eta(U_X(t,\theta,c))\le2b. The limiting activity estimate is (6): η(H)≤2b+3(13C(1+3d)36d−1)θd−1.\eta(H)\le2b+3\bigl(13C(1+3^d)36^{d-1}\bigr)\theta^{d-1}.(6)✦

Let 𝒥\mathscr J be countable and let pn,j∈𝒫p_{n,j}\in\mathcal P for every n,jn,j. There is a strictly increasing ϕ:ℕ→ℕ\phi:\mathbb N\to\mathbb N and a common parameter limit satisfying (7): pϕ(n),j→(bj,tj,θj,cj)∈𝒫(j∈𝒥).p_{\phi(n),j}\longrightarrow(b_j,t_j,\theta_j,c_j)\in\mathcal P \qquad(j\in\mathscr J).(7)✦

Let fnf_n and gg be real functions on an index set, and let TT be a finite subset of that set. Suppose fn(j)→g(j)f_n(j)\to g(j) for every index jj. For a real bound BB, if ∑j∈Tfn(j)≤B\sum_{j\in T}f_n(j)\le B eventually, the limiting budget satisfies (8): ∑j∈Tg(j)≤B.\sum_{j\in T}g(j)\le B.(8)✦

Let FnF_n be generic batch configurations with joint convergence (3) and Rn→∞R_n\to\infty, and suppose each batch system embeds into its canonical rank sweep. Assume that there is C≥0C\ge0 such that 4608an2Nn4E*,n3≤Cfor every n,13824Nn3E*,n3→0,E*,n=Rnan2.\frac{4608a_n^2N_n^4}{E_{*,n}^3}\le C\quad\text{for every }n, \qquad \frac{13824N_n^3}{E_{*,n}^3}\longrightarrow0, \qquad E_{*,n}=R_na_n^2.(9) Fix a window KK. Choose physical maps ψn(p)=rnEn−1(p−on)\psi_n(p)=r_nE_n^{-1}(p-o_n) with rn>0r_n>0. For every nn, suppose there are An,Bn>0A_n,B_n>0 and ℓn<un\ell_n<u_n such that En(1,0)=An(1,ℓn)E_n(1,0)=A_n(1,\ell_n) and En(0,1)=Bn(1,un)E_n(0,1)=B_n(1,u_n), and every sample slope of the KK window lies in [ℓn,un][\ell_n,u_n].

Suppose the enhanced point, vertex and edge measures converge weakly to μˆ,μˆJ,λˆJ\widehat\mu,\widehat\mu_J,\widehat\lambda_J, respectively, for every JJ, and their label projections are μ,μJ,λJ\mu,\mu_J,\lambda_J. Let G⊆ℒG\subseteq\mathcal L be measurable and let π:ℒ×Q→ℒ\pi:\mathcal L\times Q\to\mathcal L be label projection; set H=π−1(G)∩supp⁡μˆH=\pi^{-1}(G)\cap\operatorname{supp}\widehat\mu. Assume ∥ρ(z)∥≤1\|\rho(z)\|\le1 and enc⁡(ρ(z))=q(z)\operatorname{enc}(\rho(z))=q(z) on HH, and that for some D≥0D\ge0, ρ*(μˆ|H)(B¯(x,s))≤Ds32/17\rho_*(\widehat\mu|_H)(\overline B(x,s))\le Ds^{32/17} for every planar xx and s>0s>0. Under these hypotheses and (9), the selected window satisfies (10): λK(G)=0.\lambda_K(G)=0.(10)✦

In the paper, the result reads: for the enhanced measures of §6, a fixed dyadic window K0K_0, and a measurable set 𝒢\mathcal G of enhanced labels, assume μˆJ≤μˆ\widehat\mu_J\le\widehat\mu, λˆK0(𝒢)>0\widehat\lambda_{K_0}(\mathcal G)>0, the local cubic mass bound and activity-tail estimate, and boundedly supported β=ρ*(μˆ|𝒢)\beta=\rho_*(\widehat\mu|_{\mathcal G}) with β(B(x,r))≤CGrd\beta(B(x,r))\le C_Gr^d for 0<r<10<r<1 and d=32/17>7/4d=32/17>7/4; choose a bounded open parallelogram BB containing its support, whose side covectors are positive on the fixed pointed cone of forward transfers in K0K_0, and along one fixed matching per batch take common subsequential limits bJ=lim⁡Cn(J)/Nnb_J=\lim C_n(J)/N_n of normalized boundary counts and θJ\theta_J of angular spans, with ∑JbJ≤4\sum_Jb_J\le4 and ∑JθJ≤π\sum_J\theta_J\le\pi on each dyadic partition; then μˆJ(𝒢)≤2bJ+CGθJd−1\widehat\mu_J(\mathcal G)\le2b_J+C_G\theta_J^{d-1} for every dyadic J⊆K0J\subseteq K_0.

What Is Checked

The moving-strip and activity statements concern finite Borel measures on X×[0,1]2X\times[0,1]^2; the zero-mass statement concerns generic batch configurations embedded in canonical sweeps, jointly convergent label measures, and spatial enhancements satisfying the frame and square-mass hypotheses above.

The exchange drawing has rightward edges. Every exchange edge admits a token trace.✦