OctoInformBeta

OpenAI preprint, September 2026 · §4. Small cells and fiber mass · §4

Lemma 4.1 [Small fiber slices]: The mass T(λ)T(\lambda) of fiber slices below proportion λ\lambda, bounded through the overlap scale W2W_2 and the potential scale AA

A bound on the cell mass of small fiber slices.

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 configuration PP of nn points in the plane. The points of PP at one distance from xx lie on a circle about xx. Read as complex numbers, with Z1(y)=yZ_1(y)=y and Z2(y)=y‾Z_2(y)=\bar y, they are the points yy at which (Z1(y)−Z1(x))(Z2(y)−Z2(x))=|y−x|2(Z_1(y)-Z_1(x))(Z_2(y)-Z_2(x))=|y-x|^2 takes one value: a fiber of this product with source xx.

Spread the points of PP over the cells of a hierarchy, the nested cells given by an ultrametric depth, at every scale. A cell containing the source may hold a large share of a fiber or a small one. The theorem of this page bounds the total weight of the small shares, those below a proportion λ\lambda, summed over all cells with the mass each cell carries.

In the library, Z1,Z2:P→KZ_1,Z_2\colon P\to K are any two injective maps into a number field KK containing a square root of −1-1, and the fibers are those of t(x,y)=(Z1(y)−Z1(x))(Z2(y)−Z2(x))t(x,y)=\bigl(Z_1(y)-Z_1(x)\bigr)\bigl(Z_2(y)-Z_2(x)\bigr)(1) in yy for a fixed xx. The hierarchies come from the places of KK: a random dyadic grid at each archimedean place and the absolute value at each finite place.

Depths and Cells

An ultrametric depth dd on PP is a symmetric real function on ordered pairs of points of PP that satisfies the ultrametric inequality min⁡(d(x,y),d(y,z))≤d(x,z)for pairwise distinct x,y,z∈P.\min\bigl(d(x,y),d(y,z)\bigr)\le d(x,z)\qquad\text{for pairwise distinct }x,y,z\in P.(2)✦ The cell Cr(x)C_r(x) of a point xx at level rr is the set of the points yy of PP with y=xy=x or d(x,y)≥rd(x,y)\ge r✦, and the cells at level rr are the distinct sets Cr(x)C_r(x).✦ A cell at level rr is giant when it holds more than half of the points of PP.✦ The transformed family 𝒯r\mathcal T_r at level rr is the multiset of subsets of PP made of the cells at level rr that are not giant, each once, and the complements of the giant cells✦; Nr(C)N_r(C) is the number of times a subset CC occurs in it.✦

The cell mass of a subset CC of PP is md(C)=∫ℝNr(C)drif |C|≥2,md(C)=0if |C|<2.m_d(C)=\int_{\mathbb R}N_r(C)\,dr\quad\text{if }|C|\ge2,\qquad m_d(C)=0\quad\text{if }|C|<2.(3)✦ The potential of a point xx of PP is Sd(x)=∫ℝ𝟏[some cell at level r is giant and none of the giant cells contains x]dr.S_d(x)=\int_{\mathbb R}\mathbf 1\bigl[\text{some cell at level $r$ is giant and none of the giant cells contains $x$}\bigr]\,dr.(4)✦ The overlap length of two points xx, yy of PP is Ld(x,y)=∫ℝ#{C∈𝒯r:x∈C,y∈C}dr,L_d(x,y)=\int_{\mathbb R}\#\{C\in\mathcal T_r:\ x\in C,\ y\in C\}\,dr,(5)✦ the members of 𝒯r\mathcal T_r that contain both points being counted with multiplicity.

The Hierarchies of a Number Field

Fix a coordinate ZjZ_j, j=1,2j=1,2. At a finite place vv of KK the raw depth of ZjZ_j is the depth iji_j on PP given by ij(x,y)=−log⁡|Zj(y)−Zj(x)|vi_j(x,y)=-\log|Z_j(y)-Z_j(x)|_v for distinct xx, yy and ij(x,x)=0i_j(x,x)=0.✦ The bad places of ZjZ_j are the finitely many finite places at which some difference Zj(y)−Zj(x)Z_j(y)-Z_j(x) of distinct points is not a unit.✦

At an archimedean place the hierarchy is a random shifted dyadic grid, rotated and rescaled at random. A sample ω\omega consists of a phase (ξ,ϕ)(\xi,\phi) and two axis sequences ω(1)\omega^{(1)}, ω(2)\omega^{(2)}✦, all independent: ξ\xi is uniform on (0,log⁡2](0,\log 2], ϕ\phi is uniform on (0,2π](0,2\pi]✦, and each axis sequence has independent terms✦, the term of index 00 uniform on (0,1](0,1] and the term of index i+1i+1 equal to 00 or 2i2^i with probability 12\frac12 each.✦ The mean over samples is written 𝔼ω\mathbb E_\omega.

For an integer mm and an axis k=1,2k=1,2, the offset σm(ω(k))\sigma_m(\omega^{(k)}) is the sum of the terms of ω(k)\omega^{(k)} of index 00 to mm, or its term of index 00 alone when m<0m<0.✦ Two points uu, u′u' of ℝ2\mathbb R^2 lie in one cell at scale mm when on both axes they fall in the same interval of length 2m2^m of the grid with that offset, ⌊uk−σm(ω(k))2m⌋=⌊uk′−σm(ω(k))2m⌋,k=1,2,\Bigl\lfloor\frac{u_k-\sigma_m(\omega^{(k)})}{2^m}\Bigr\rfloor =\Bigl\lfloor\frac{u'_k-\sigma_m(\omega^{(k)})}{2^m}\Bigr\rfloor,\qquad k=1,2,(6)✦ and their dyadic depth is gω(u,u′)=−m*log⁡2g_\omega(u,u')=-m_*\log 2, with m*m_* the infimum of the scales at which they lie in one cell.✦ The hierarchy depth of two complex numbers zz, z′z', read as points of ℝ2\mathbb R^2 through their real and imaginary parts, is δω(z,z′)=gω(e−ξeiϕz,e−ξeiϕz′)−ξ.\delta_\omega(z,z')=g_\omega\bigl(e^{-\xi}e^{i\phi}z,\,e^{-\xi}e^{i\phi}z'\bigr)-\xi.(7)✦

Let vv be an archimedean place with complex embedding ιv\iota_v. A sample ω\omega is good at vv for ZjZ_j when δω\delta_\omega satisfies the ultrametric inequality (2) on the points ιv(Zj(x))\iota_v(Z_j(x)), x∈Px\in P.✦ The grid depth of ZjZ_j at vv is the depth djd_j on PP with dj(x,y)=δω(ιv(Zj(x)),ιv(Zj(y)))d_j(x,y)=\delta_\omega\bigl(\iota_v(Z_j(x)),\iota_v(Z_j(y))\bigr) for distinct xx, yy when ω\omega is good, and dj(x,y)=0d_j(x,y)=0 otherwise.✦

Each archimedean place vv carries the weight mult⁡(v)/[K:ℚ]\operatorname{mult}(v)/[K:\mathbb Q], where mult⁡(v)\operatorname{mult}(v) is 11 at a real place and 22 at a complex one, and each finite place the weight 1/[K:ℚ]1/[K:\mathbb Q]. The cell mass of ZjZ_j is mj(C)=∑v∣∞mult⁡(v)[K:ℚ]𝔼ωmdj(C)+1[K:ℚ]∑v badmij(C),m_j(C)=\sum_{v\mid\infty}\frac{\operatorname{mult}(v)}{[K:\mathbb Q]}\, \mathbb E_\omega\,m_{d_j}(C)+\frac1{[K:\mathbb Q]}\sum_{v\text{ bad}}m_{i_j}(C),(8)✦ The potential Sj(x)S_j(x) of ZjZ_j at a point xx of PP is the same weighted sum with Sdj(x)S_{d_j}(x) and Sij(x)S_{i_j}(x) in place of mdj(C)m_{d_j}(C) and mij(C)m_{i_j}(C). The total cell mass is m(C)=m1(C)+m2(C)m(C)=m_1(C)+m_2(C)✦, the total potential is S*(x)=S1(x)+S2(x)S^*(x)=S_1(x)+S_2(x)✦, and for a function ff on subsets of PP, I(f)=∑C⊆Pm(C)f(C)I(f)=\sum_{C\subseteq P}m(C)\,f(C).✦

The uniform law π\pi on distinct ordered pairs of PP gives each pair (x,y)(x,y) with x≠yx\ne y the mass 1/(n(n−1))1/(n(n-1))✦, and π(C×C)\pi(C\times C) is the mass it gives to the pairs inside a subset CC.✦ With α\alpha the uniform law on PP, the overlap scale W2W_2 and the potential scale AA are W2=∑C⊆Pm(C)π(C×C),A=𝔼αS*=1n∑x∈PS*(x).W_2=\sum_{C\subseteq P}m(C)\,\pi(C\times C),\qquad A=\mathbb E_\alpha S^*=\frac1n\sum_{x\in P}S^*(x).(9)

Fiber Families and Slices

A weighted fiber family on PP is a finite family of triples (wi,pi,xi)(w_i,p_i,x_i), each made of a weight wi≥0w_i\ge0, a probability law pip_i on PP and a source xi∈Px_i\in P with pi(xi)=0p_i(x_i)=0, the weights summing to 11. The edge law of a weighted fiber family (wi,pi,xi)(w_i,p_i,x_i) on PP is the function qq on pairs of points of PP given by q(a,b)=∑i:xi=awipi(b)(a,b∈P),q(a,b)=\sum_{i:\,x_i=a}w_i\,p_i(b)\qquad(a,b\in P),(10) A weighted fiber family (wi,pi,xi)(w_i,p_i,x_i) on PP lies on fibers of tt when each index ii has a value ti∈Kt_i\in K with t(xi,y)=tit(x_i,y)=t_i for every yy such that pi(y)>0p_i(y)>0.

For a subset CC of PP, pi(C)=∑y∈Cpi(y)p_i(C)=\sum_{y\in C}p_i(y) is the mass of CC under pip_i.✦ The λ\lambda-slice of the family is the function on subsets of PP that keeps, from the laws with source in CC, the mass they give to CC when it is less than λ\lambda: Tλ(C)=∑i:xi∈C,pi(C)<λwipi(C).T_\lambda(C)=\sum_{i:\;x_i\in C,\ p_i(C)<\lambda}w_i\,p_i(C).(11)✦ The slice mass is T(λ)=I(Tλ)T(\lambda)=I(T_\lambda).

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.✦

For a set CC of vertices and a level ℓ\ell, a fiber ee is active in CC when xe∈Cx_e\in C and more than ℓ\ell of its targets lie in CC✦, and the edges at level ℓ\ell in CC are the pairs (xe,y)(x_e,y) with ee active and y∈Fe∩Cy\in F_e\cap C.✦ At level −1-1 in VV every fiber is active, and these edges form the edge set EE of GG, the pairs (xe,y)(x_e,y) with y∈Fey\in F_e.✦ The degree of a vertex vv at level ℓ\ell in CC is the number of edges at level ℓ\ell in CC leaving vv plus the number entering vv; the degree of vv in GG is its degree at level −1-1 in VV, that is, the number of edges of EE leaving it plus the number entering it.✦ The fiber weight of ee is we=|Fe|/|E|w_e=|F_e|/|E|, the mass that the uniform law qq on EE gives to the edges of ee, and μe\mu_e is the uniform law on FeF_e.

For a proportion β\beta and a set DD of vertices, the heavy count Eβ(D)E_\beta(D) is the number of edges inside DD carried by fibers with source in DD that keep at least the proportion β\beta of their targets in DD: Eβ(D)=∑e:xe∈D,|Fe∩D|≥β|Fe||Fe∩D|.E_\beta(D)=\sum_{e:\;x_e\in D,\ |F_e\cap D|\ge\beta|F_e|}|F_e\cap D|.(12)✦

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\ge2, (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),(13)✦ every vertex has degree at least 2(θ−η)(N−1)2(\theta-\eta)(N-1), every fiber has more than NγN^\gamma targets, and every set DD of vertices has E1/log⁡N(D)≤(θ+η)|D|(|D|−1)E_{1/\log N}(D)\le(\theta+\eta)|D|(|D|-1), and E1/log⁡N(D)≤η|D|(|D|−1)E_{1/\log N}(D)\le\eta|D|(|D|-1) when |D|≤N1−c|D|\le N^{1-c}.

The Claim

Theorem 1. Let PP be a configuration of n≥2n\ge2 points with injective coordinates Z1Z_1, Z2Z_2 in KK, and let 0≤λ≤120\le\lambda\le\frac12, 0<τ≤10<\tau\le1 with τn≥2\tau n\ge2, and H≥0H\ge0.

(a) Let a weighted fiber family (wi,pi,xi)(w_i,p_i,x_i) on PP have edge law q(a,b)≤B/(n(n−1))q(a,b)\le B/(n(n-1)) for all distinct aa, bb, where B≥0B\ge0. For any two ultrametric depths dd, d′d' on PP and any real numbers hih_i, one for each index, ∑C⊆Pmd(C)Tλ(C)≤2λτ2∑C⊆Pmd(C)π(C×C)+Bτn∑y∈P(Sd(y)+Sd′(y))+2λ∑iwi∑y∈Ppi(y)Ld′(xi,y)+2λH+∑iwi∑y∈Ppi(y)max⁡(|d(xi,y)+d′(xi,y)−hi|−H,0).\begin{split} \sum_{C\subseteq P}m_d(C)\,T_\lambda(C)\le{}&\frac{2\lambda}{\tau^2}\sum_{C\subseteq P}m_d(C)\,\pi(C\times C) +\frac{B\tau}{n}\sum_{y\in P}\bigl(S_d(y)+S_{d'}(y)\bigr)\\ &+2\lambda\sum_iw_i\sum_{y\in P}p_i(y)\,L_{d'}(x_i,y)+2\lambda H\\ &+\sum_iw_i\sum_{y\in P}p_i(y)\max\bigl(|d(x_i,y)+d'(x_i,y)-h_i|-H,\,0\bigr). \end{split}(14)✦

(b) If the family of (a) lies on fibers of tt, its slice mass satisfies T(λ)≤(2λτ2+2λB)W2+2BτA+4λH+32e−H/2.T(\lambda)\le\Bigl(\frac{2\lambda}{\tau^2}+2\lambda B\Bigr)W_2+2B\tau A+4\lambda H+32e^{-H/2}.(15)✦

(c) Let GG be a prepared graph on PP with density θ\theta, exponents γ\gamma, cc and error η<θ\eta<\theta, each of whose fibers ee has a value te∈Kt_e\in K with t(xe,y)=tet(x_e,y)=t_e for all y∈Fey\in F_e. The slice mass T(λ)T(\lambda) built from the fiber weights wew_e, the uniform laws μe\mu_e on FeF_e and the sources xex_e satisfies T(λ)≤(2λτ2+2λθ−η)W2+2τθ−ηA+4λH+32e−H/2.T(\lambda)\le\Bigl(\frac{2\lambda}{\tau^2}+\frac{2\lambda}{\theta-\eta}\Bigr)W_2 +\frac{2\tau}{\theta-\eta}\,A+4\lambda H+32e^{-H/2}.(16)✦

In the paper, the result reads: for the prepared graph and its hierarchies, if 0<λ<1/20<\lambda<1/2, 2/n≤τ≤12/n\le\tau\le1 and H≥1H\ge1, then T(λ)≪(λ/τ2+λ)W2+τA+λH+e−b1HT(\lambda)\ll(\lambda/\tau^2+\lambda)W_2+\tau A+\lambda H+e^{-b_1H}, with b1>0b_1>0 absolute and the implied constant depending only on B0B_0 and on the constants of the paper’s (3.4).

What Is Checked

The claim is stated over a configuration PP of at least two points with injective coordinates Z1Z_1, Z2Z_2 in a number field KK; over ultrametric depths with their cells, cell masses, potentials and overlap lengths; over the raw depths at the bad finite places and the grid depths at the archimedean places, weighted and combined into the total cell mass mm, the total potential S*S^*, the overlap scale W2W_2 and the potential scale AA; over weighted fiber families with their edge law and λ\lambda-slices; and over prepared graphs with their fiber weights and the uniform laws on their target sets.

The library shows that the raw depth iji_j at every finite place, and the grid depth djd_j at every archimedean place for every sample, are ultrametric depths on PP.