OctoInformBeta

OpenAI preprint, September 2026 · §6. The case of unbounded overlap scale · §6

§6 (proof): the unbounded case is impossible: Along a sequence of growing size with errors ηj→0\eta_j\to 0 and c>0c>0; for density θ\theta the potential scale is eventually at most (64+8/θ)(W2+1)(64+8/\theta)(W_2+1), with W2W_2 the overlap scale

Prepared arithmetic models cannot have overlap scales tending to infinity.

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

Pin a point xx of a finite configuration PP in the plane. The other points of PP lie on circles centred at xx, one circle for each distance from xx, and two circles around the same pin are disjoint. Writing a point as u+ibu+ib and setting Z1=u+ibZ_1=u+ib, Z2=u−ibZ_2=u-ib, the product t(x,y)=(Z1(y)−Z1(x))(Z2(y)−Z2(x))=|y−x|2t(x,y)=\bigl(Z_1(y)-Z_1(x)\bigr)\bigl(Z_2(y)-Z_2(x)\bigr)=|y-x|^2(1) takes a single value on each circle around xx.

An arithmetic model keeps only this structure. The pins and their circles become the fibers of a graph on a finite vertex set; the coordinates become two injective maps Z1Z_1, Z2Z_2 from the vertices into a number field KK containing a square root of −1-1; and tt is required to be constant on each fiber. Closeness of two vertices is then read at every place of KK: at an archimedean place through a randomly rotated and shifted dyadic grid, at a finite place through the vv-adic absolute value of Zj(y)−Zj(x)Z_j(y)-Z_j(x). Each reading is an ultrametric hierarchy of cells, and the cells of all of them together give two numbers, the overlap scale W2W_2 and the potential scale AA. The case at hand is a sequence of such models of growing size along which W2W_2 tends to infinity.

Prepared Graphs

A fiber system GG on a finite vertex set VV, with fibers indexed by a finite set II, assigns to each fiber ee a source xe∈Vx_e\in V and a target set Fe⊆VF_e\subseteq V with xe∉Fex_e\notin F_e, such that two distinct fibers with the same source have disjoint target sets.✦ The edge set EE of a fiber system GG is the set of pairs (xe,y)(x_e,y) with e∈Ie\in I and y∈Fey\in F_e; it has |E||E| edges.

For a set CC of vertices and a level τ\tau, the size of a fiber ee in CC is |Fe∩C||F_e\cap C|✦, and ee is active when xe∈Cx_e\in C and its size in CC exceeds τ\tau.✦ The graph of GG at level τ\tau on CC has an edge from xx to yy when some active fiber has source xx and y∈Fe∩Cy\in F_e\cap C✦, and the degree of a vertex in it is its out-degree plus its in-degree.✦ At level −1-1 on all of VV every fiber is active and the edge set is EE; the degree of a vertex of VV in GG, with no level named, is its degree in the graph of GG at level −1-1 on VV.

For a fiber system GG, a real λ\lambda and a set DD of vertices, the heavy count hλ(D)h_\lambda(D) of DD is the total number of targets inside DD of the fibers with source in DD that keep at least a fraction λ\lambda of their targets in DD: hλ(D)=∑e:xe∈D|Fe∩D|≥λ|Fe||Fe∩D|h_\lambda(D)=\sum_{\substack{e:\ x_e\in D\\ |F_e\cap D|\ge\lambda|F_e|}}|F_e\cap D|(2)✦

A fiber system GG on VV, with N=|V|N=|V|, is a prepared graph with density θ\theta, exponents γ\gamma, cc and error η\eta when N≥2N\ge 2, every vertex has degree at least 2(θ−η)(N−1)2(\theta-\eta)(N-1), every fiber has |Fe|>Nγ|F_e|>N^{\gamma}, (θ−η)N(N−1)≤|E|≤(θ+η/8)N(N−1),(\theta-\eta)N(N-1)\le|E|\le(\theta+\eta/8)N(N-1),(3) and every set DD of vertices satisfies, with λ=1/log⁡N\lambda=1/\log N, hλ(D)≤(θ+η)|D|(|D|−1),hλ(D)≤η|D|(|D|−1) if |D|≤N1−c.h_\lambda(D)\le(\theta+\eta)|D|(|D|-1),\qquad h_\lambda(D)\le\eta|D|(|D|-1)\ \text{ if } |D|\le N^{1-c}.(4)✦

A prepared arithmetic model is a prepared graph GG on VV together with a number field KK and two injective maps Z1,Z2:V→KZ_1,Z_2\colon V\to K such that, for every fiber ee, the product t(xe,y)=(Z1(y)−Z1(xe))(Z2(y)−Z2(xe))t(x_e,y)=\bigl(Z_1(y)-Z_1(x_e)\bigr)\bigl(Z_2(y)-Z_2(x_e)\bigr) takes one value te∈Kt_e\in K for all y∈Fey\in F_e.

Ultrametric Hierarchies

An ultrametric depth dd on a finite set PP of nn points is a symmetric real function on pairs of points of PP such that d(x,z)≥min⁡(d(x,y),d(y,z))d(x,z)\ge\min\bigl(d(x,y),d(y,z)\bigr) for pairwise distinct xx, yy, zz.✦

Two points xx, yy of PP share a cell at level r∈ℝr\in\mathbb{R} when they are equal or d(x,y)≥rd(x,y)\ge r.✦ The cell Cr(x)C_r(x) is the set of points sharing a cell with xx at level rr✦, and 𝒞r\mathcal{C}_r is the set of all cells at level rr✦. A giant at level rr is a cell C∈𝒞rC\in\mathcal{C}_r with 2|C|>n2|C|>n; the giants form the set 𝒢r\mathcal{G}_r✦, and GrG_r is their union✦. The transformed family 𝒯r\mathcal{T}_r is the multiset made of the cells C∈𝒞rC\in\mathcal{C}_r with 2|C|≤n2|C|\le n and the complements P\CP\setminus C of the giants, a set occurring in both parts counted twice✦, and 𝒯r(C)\mathcal{T}_r(C) is the number of times CC occurs in it✦.

For an ultrametric depth dd on PP, the cell mass md(C)m_d(C) of a subset CC of PP with |C|≥2|C|\ge 2 and the potential Sd(x)S_d(x) of a point xx of PP are the real numbers md(C)=∫ℝ𝒯r(C)dr,Sd(x)=∫ℝ𝟏[𝒢r≠∅,x∉Gr]dr,m_d(C)=\int_{\mathbb{R}}\mathcal{T}_r(C)\,dr,\qquad S_d(x)=\int_{\mathbb{R}}\mathbf{1}\bigl[\mathcal{G}_r\ne\emptyset,\ x\notin G_r\bigr]\,dr,(5) and a subset with fewer than two points has cell mass 00.

Depths at the Places of a Number Field

A phase and grid sample ω\omega is a phase (ξ,φ)∈ℝ2(\xi,\varphi)\in\mathbb{R}^2, made of a scale shift ξ\xi and a rotation angle φ\varphi, together with a pair of axis sequences (ω1,i)i≥0(\omega_{1,i})_{i\ge0} and (ω2,i)i≥0(\omega_{2,i})_{i\ge0} of real numbers.✦ The axis law is the law of a sequence of independent real numbers whose term of index 00 is uniform on (0,1](0,1] and whose term of index i+1i+1 equals 00 or 2i2^i with probability 1/21/2 each.✦ The full sample law on phase and grid samples makes ξ\xi uniform on (0,log⁡2](0,\log 2], φ\varphi uniform on (0,2π](0,2\pi] and the two sequences independent copies of the axis law, all four independent.✦

For k=1,2k=1,2 and an integer mm, the offset of the sequence ωk\omega_k at scale mm is the partial sum sk,m=ωk,0+…+ωk,ms_{k,m}=\omega_{k,0}+\dots+\omega_{k,m} when m≥0m\ge0 and sk,m=ωk,0s_{k,m}=\omega_{k,0} when m<0m<0.✦ Two points xx, yy of ℝ2\mathbb{R}^2 lie in the same cell at scale mm when, in each coordinate kk, they have the same index ⌊(xk−sk,m)/2m⌋\lfloor(x_k-s_{k,m})/2^m\rfloor.✦ The grid depth gω(x,y)g_\omega(x,y) is −mlog⁡2-m\log 2 for the infimum mm of the scales at which xx and yy share a cell✦, and the hierarchy depth of two complex numbers is δω(z,w)=gω(e−ξeiφz,e−ξeiφw)−ξ,\delta_\omega(z,w)=g_\omega\bigl(e^{-\xi}e^{i\varphi}z,\ e^{-\xi}e^{i\varphi}w\bigr)-\xi,(6)✦ the rotated and rescaled points read as points of ℝ2\mathbb{R}^2 through their real and imaginary parts.

Let ZZ be an injective map from a finite set PP into a number field KK. A sample ω\omega is good at an archimedean place vv when δω\delta_\omega, taken at the images v(Z(x))v(Z(x)) under the complex embedding at vv, satisfies the ultrametric inequality for pairwise distinct points of PP.✦ The archimedean depth dv,ωd_{v,\omega} is δω(v(Z(x)),v(Z(y)))\delta_\omega\bigl(v(Z(x)),v(Z(y))\bigr) at distinct points xx, yy when ω\omega is good at vv, and 00 otherwise.✦ At a finite place vv the finite place depth dvd_v vanishes on the diagonal and is the raw depth dv(x,y)=−log⁡|Z(x)−Z(y)|vd_v(x,y)=-\log|Z(x)-Z(y)|_v at distinct points.✦ The bad places of ZZ are the finite places at which the difference Z(x)−Z(y)Z(x)-Z(y) of some pair of distinct points is not a unit.✦

Each archimedean place carries the weight σv=mult⁡(v)/[K:ℚ]\sigma_v=\operatorname{mult}(v)/[K:\mathbb{Q}], where mult⁡(v)\operatorname{mult}(v) is 11 for a real place and 22 for a complex one.✦ With 𝔼ω\mathbb{E}_\omega the mean under the full sample law, the global cell mass of a subset CC of PP is mZ(C)=∑v∣∞σv𝔼ωmdv,ω(C)+1[K:ℚ]∑vbadmdv(C),m_Z(C)=\sum_{v\mid\infty}\sigma_v\,\mathbb{E}_\omega\,m_{d_{v,\omega}}(C) +\frac{1}{[K:\mathbb{Q}]}\sum_{v\ \text{bad}}m_{d_v}(C),(7)✦ and the potential SZ(x)S_Z(x) of a point is the same sum with Sd(x)S_d(x) in place of md(C)m_d(C).✦ For two injective maps Z1Z_1, Z2Z_2 of PP into KK, the total cell mass and the total potential collect both coordinates: m*(C)=mZ1(C)+mZ2(C),S*(x)=SZ1(x)+SZ2(x).m^*(C)=m_{Z_1}(C)+m_{Z_2}(C),\qquad S^*(x)=S_{Z_1}(x)+S_{Z_2}(x).(8)

Overlap and Potential Scales

On a finite set BB, α\alpha is the uniform law on BB✦ and π\pi the uniform law on its distinct ordered pairs, which gives each pair of distinct points of BB the mass 1/(|B|(|B|−1))1/(|B|(|B|-1))✦; 𝔼α\mathbb{E}_\alpha and 𝔼π\mathbb{E}_\pi are the means under them, and, for a subset CC of BB, π(C×C)\pi(C\times C) is the probability that a π\pi-random pair lies in C×CC\times C✦. For a function mm on the subsets of PP and a real function ff on the subsets of PP, the weighted integral of ff is Im(f)=∑C⊆Pm(C)f(C)I_m(f)=\sum_{C\subseteq P}m(C)f(C).✦ The weighted overlap Lm(x,y)L_m(x,y) of two points xx, yy of PP is Lm(x,y)=Im(𝟏[x∈C,y∈C])=∑C∋x,ym(C).L_m(x,y)=I_m\bigl(\mathbf{1}[x\in C,\ y\in C]\bigr)=\sum_{C\ni x,y}m(C).(9)✦

For two injective maps Z1Z_1, Z2Z_2 of PP into KK, with α\alpha and π\pi the uniform laws on all of PP and L*=Lm*L^*=L_{m^*}, the overlap scale and the potential scale are W2=Im*(π(C×C))=𝔼πL*,A=𝔼αS*.W_2=I_{m^*}\bigl(\pi(C\times C)\bigr)=\mathbb{E}_\pi L^*,\qquad A=\mathbb{E}_\alpha S^*.(10)

Cut Distances

A family ss of subsets of a finite set PP is laminar when any two of its members are nested or disjoint.✦

Let ss be a finite family of subsets of PP, whose members are the cuts, each member CC having a real weight l(C)l(C). A real weighting of PP is a real function on PP. The cut distance ds(x,y)d_s(x,y) of two points xx, yy of PP and the cut cost cost⁡s(p,p′)\operatorname{cost}_s(p,p') of two real weightings pp, p′p' of PP are ds(x,y)=∑C∈sl(C)|𝟏C(x)−𝟏C(y)|,cost⁡s(p,p′)=∑C∈sl(C)|p(C)−p′(C)|,d_s(x,y)=\sum_{C\in s}l(C)\,\bigl|\mathbf{1}_C(x)-\mathbf{1}_C(y)\bigr|,\qquad \operatorname{cost}_s(p,p')=\sum_{C\in s}l(C)\,\bigl|p(C)-p'(C)\bigr|,(11) where 𝟏C\mathbf{1}_C is the indicator of CC and p(C)=∑x∈Cp(x)p(C)=\sum_{x\in C}p(x) is the total weight of CC. For a weighting aa of PP, the double mean fluctuation of dsd_s at a point yy is Δa(y)=∑x′a(x′)∑xa(x)|ds(x,y)−ds(x′,y)|.\Delta_a(y)=\sum_{x'}a(x')\sum_{x}a(x)\,\bigl|d_s(x,y)-d_s(x',y)\bigr|.(12)✦

The Claim

Theorem 1. (a) Let PP be a finite set of at least three points, mm a nonnegative function on the subsets of PP, bb a real number and SS a real function on PP such that Lm(x,y)=b+S(x)+S(y)L_m(x,y)=b+S(x)+S(y) for all distinct xx, yy. Put A=𝔼αSA=\mathbb{E}_\alpha S and W2=Im(π(C×C))W_2=I_m\bigl(\pi(C\times C)\bigr) for the uniform laws on PP. Then every point xx satisfies S(x)≥A−2W2.S(x)\ge A-2W_2.(13)✦

(b) Let θ>0\theta>0, 0<γ<10<\gamma<1 and c<γc<\gamma. For each jj let GjG_j, with coordinates Z1,jZ_{1,j}, Z2,jZ_{2,j} into a number field KjK_j, be a prepared arithmetic model on a vertex set VjV_j with density θ\theta, exponents γ\gamma, cc and error ηj\eta_j, where |Vj|→∞|V_j|\to\infty, ηj→0\eta_j\to0 and 0≤ηj≤θ/20\le\eta_j\le\theta/2, and let AjA_j and W2,jW_{2,j} be its potential and overlap scales. Then for all sufficiently large jj Aj≤(64+8θ)(W2,j+1).A_j\le\Bigl(64+\frac{8}{\theta}\Bigr)\bigl(W_{2,j}+1\bigr).(14)✦

(c) Let ss be a laminar family of subsets of a finite set PP with weights l(C)≥0l(C)\ge0 on its members, let aa and ν\nu be probability vectors on PP, let ρy\rho_y be a probability vector on PP for each y∈Py\in P, and let RR be a real function on PP. Then ∑ya(y)Δa(y)≤2∑yν(y)∑xρy(x)|ds(x,y)−R(y)|+2∑yν(y)cost⁡s(ρy,a)+2cost⁡s(ν,a).\sum_y a(y)\,\Delta_a(y)\le 2\sum_y\nu(y)\sum_x\rho_y(x)\,\bigl|d_s(x,y)-R(y)\bigr| +2\sum_y\nu(y)\operatorname{cost}_s(\rho_y,a)+2\operatorname{cost}_s(\nu,a).(15)✦

(d) Under the hypotheses of (b), if moreover c>0c>0, the overlap scales do not tend to infinity: W2,j↛∞(j→∞).W_{2,j}\not\to\infty\qquad(j\to\infty).(16)✦

In the paper, the result reads: W2≥c1WW_2\ge c_1W with W=W2+AW=W_2+A and a constant c1c_1, this being inequality (6.1), and along a subsequence with W→∞W\to\infty the variance bounds and the tree transport contradict Lemma 6.1 of the paper.

What Is Checked

The claims are stated over fiber systems and prepared graphs with their edges, degrees and heavy counts; ultrametric depths with their cells, giants, transformed families, cell masses and potentials; the randomly rotated and shifted dyadic grid and its sample law; the archimedean and finite place depths of an injective map into a number field, with its bad places and the weights σv\sigma_v; the total cell mass and total potential, with the overlap and potential scales built from them; and cut distances, cut costs and double mean fluctuations over a laminar family.

The library proves that the finite place depth and, for every phase and grid sample, the archimedean depth are ultrametric depths, and that an injective map of a finite set into a number field has only finitely many bad places.