OctoInformBeta

OpenAI preprint, September 2026 · §3. A metric and volume bound for connected supports

Ball volume and the high-dimensional margin: The degree bound gives an explicit dimension threshold for a strict radius margin

Transportation ball volumes admit factorial bounds.

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

The geometric quantity is the area of a transportation ball in a simplicial realization, with marked subsets varying over a parameter space. Let Γ\Gamma be a connected simple graph on a vertex set YY, with graph metric dΓd_\Gamma given by shortest-path distance, and let KK be an abstract simplicial complex on YY. The real vector space E=ℝ(Y)E=\mathbb R^{(Y)} of finitely supported vertex vectors has the norm induced by its coordinate-preserving embedding into ℓ2(Y)\ell^2(Y). For a face σ\sigma of KK, write [σ][\sigma] for the convex hull of its standard vertices {ey:y∈σ}\{e_y:y\in\sigma\} and σ∘\sigma^\circ for its intrinsic interior; the realization |K||K| is the union of these convex hulls. For a natural number nn, the set KnK_n consists of the faces with n+1n+1 vertices.✦

A finitely supported probability measure pp on YY is a function of finite support with nonnegative real coefficients summing to one.✦ For finitely supported probability measures pp and qq on YY and a real function ff on YY, the expectation and expectation difference are defined in (1): 𝔼p[f]=∑y∈Yp(y)f(y),Δp,q(f)=𝔼p[f]−𝔼q[f].\mathbb E_p[f]=\sum_{y\in Y}p(y)f(y),\qquad \Delta_{p,q}(f)=\mathbb E_p[f]-\mathbb E_q[f].(1) The dual transportation distance W1W_1 and its normalization ρn\rho_n are given by (2), with Lipschitz constants taken for dΓd_\Gamma: W1(p,q)=supLip⁡(f)≤1Δp,q(f),ρn(p,q)=W1(p,q)n.W_1(p,q)=\sup_{\operatorname{Lip}(f)\le1}\Delta_{p,q}(f),\qquad \rho_n(p,q)=\frac{W_1(p,q)}{\sqrt n}.(2) For a realization point v∈|K|v\in|K|, its barycentric probability measure pvp_v has coefficient vector vv and support a face of KK.✦ For a probability center cc on YY and a real radius ss, the center distance dcd_c and the closed ball Ls(c)L_s(c) are defined in (3): dc(v)={ρn(c,pv),v∈|K|,0,v∉|K|,Ls(c)={v∈E:dc(v)≤s}.d_c(v)= \begin{cases}\rho_n(c,p_v),&v\in|K|,\\0,&v\notin|K|,\end{cases} \qquad L_s(c)=\{v\in E:d_c(v)\le s\}.(3)✦

Marked Subdivisions and Area

For a geometric complex LL in EE, its vertex restriction L|AL|_A to A⊆EA\subseteq E retains the faces whose vertices belong to AA; its geometric restriction L⟨A⟩L\langle A\rangle retains the faces whose entire convex hull belongs to AA. A complex LL respects a subset AA when its faces satisfy (4): [s]∩A=conv⁡(s∩A)(s a face of L).[s]\cap A=\operatorname{conv}(s\cap A)\qquad(s\text{ a face of }L).(4)✦ A simplicial subdivision TT of KK is a geometric complex with underlying space |K||K|, each of whose closed simplices lies in an original closed simplex. It respects every [σ][\sigma], and T|[σ]T|_{[\sigma]} has finitely many faces for every original face σ\sigma.

A locally constant subdivision family (Tx)x∈X(T_x)_{x\in X}, for a topological space XX, is a family of subdivisions whose restricted face sets Faces⁡(Tx|[σ])\operatorname{Faces}(T_x|_{[\sigma]}) are locally constant in xx for every original face σ\sigma.✦ A marked subset PP of a geometric complex is a subset such that each point of PP belongs to a closed simplex contained in PP.✦ A marked subdivision family (Px)(P_x) for (Tx)(T_x) consists of marked subsets of TxT_x for which Faces⁡((Tx⟨Px⟩)|[σ])\operatorname{Faces}((T_x\langle P_x\rangle)|_{[\sigma]}) is locally constant in xx for each original face σ\sigma.✦

Let a group HH act on YY and XX; its vertex action on EE is linear reindexing of coordinates.✦ An equivariant marked subdivision pair Q=(T,P)Q=(T,P) for the actions of HH on YY and XX consists of a locally constant subdivision family (Tx)x∈X(T_x)_{x\in X} and a marked subdivision family (Px)x∈X(P_x)_{x\in X} such that hh sends every face of TxT_x to a face of ThxT_{hx} and sends PxP_x into PhxP_{hx}, for every h∈Hh\in H and x∈Xx\in X.✦

Let σ∈Kn\sigma\in K_n have vertex enumeration y0,…,yny_0,\ldots,y_n. The simplex coordinates are Cσ(v)=(v(y0),…,v(yn))C_\sigma(v)=(v(y_0),\ldots,v(y_n)), and the simplex graph Γσ\Gamma_\sigma is the graph on the indices 0,…,n0,\ldots,n with an edge ijij precisely when yiy_i and yjy_j are adjacent in Γ\Gamma. A spanning tree of Γσ\Gamma_\sigma is a tree subgraph on all its vertices; write 𝒯σ\mathcal T_\sigma for their set.✦ Root at vertex 00, with rooted coordinates η0=0\eta_0=0 and ηi=ei∈ℤn\eta_i=e_i\in\mathbb Z^n for 1≤i≤n1\le i\le n. For a spanning tree T∈𝒯σT\in\mathcal T_\sigma, enumerate its edges and choose ordered endpoints (ue,ve)(u_e,v_e). The integer reduced incidence matrix CTC_T has column ηue−ηve\eta_{u_e}-\eta_{v_e} at edge ee, and CTℝC_T^{\mathbb R} is its real image. The barycentric and tree parametrizations are given in (5): β(z)=(1−∑i=1nzi,z1,…,zn),ΦT(z)=β(CTℝz),z∈ℝn.\beta(z)=\left(1-\sum_{i=1}^n z_i,z_1,\ldots,z_n\right),\qquad \Phi_T(z)=\beta(C_T^{\mathbb R}z),\quad z\in\mathbb R^n.(5) An edge enumeration of a tree on n+1n+1 vertices is a bijection from {1,…,n}\{1,\ldots,n\} to its edge set.

Let jj be a nonnegative integer, and write ℋj\mathcal H^j for Euclidean Hausdorff measure on ℝn\mathbb R^n. The tree area ajTa_j^T of a set of barycentric coordinates AA, and the simplex-specific averaged area ajσa_j^\sigma of W⊆EW\subseteq E, are defined in (6); both take values in [0,∞][0,\infty]: ajT(A)=ℋj(ΦT−1(A)),ajσ(W)=1|𝒯σ|∑T∈𝒯σajT(Cσ(W)).a_j^T(A)=\mathcal H^j(\Phi_T^{-1}(A)),\qquad a_j^\sigma(W)=\frac{1}{|\mathcal T_\sigma|} \sum_{T\in\mathcal T_\sigma}a_j^T(C_\sigma(W)).(6) For a measure μ\mu on XX and a subset B⊆XB\subseteq X, the local original simplex area AjσA_j^\sigma and the integrated ball mass MjBM_j^B are defined in (7): Ajσ(W;x)=ajσ(Px∩σ∘∩W),MjB(c,s)=∫B∑σ∈KnAjσ(Ls(c);x)dμ(x).\begin{aligned} A_j^\sigma(W;x)&=a_j^\sigma(P_x\cap\sigma^\circ\cap W),\\ M_j^B(c,s)&=\int_B\sum_{\sigma\in K_n}A_j^\sigma(L_s(c);x)\,d\mu(x). \end{aligned}(7) The sum and integral in (7) are extended nonnegative ones.

The Claim

For a natural number DD and a real number tt, the ball volume bound VD,nV_{D,n}, divided power t[n]t^{[n]}, and dimension threshold N(D)N(D) are the quantities in (8); the subscript ++ denotes the nonnegative integer part: VD,n(s)=(n+1)D⌊sn⌋++1D2nn!,t[n]=tnn!,N(D)=(4D3+D+3)2.\begin{aligned} V_{D,n}(s)&=\frac{(n+1)D^{\lfloor s\sqrt n\rfloor_++1}D^{2n}}{n!},\\ t^{[n]}&=\frac{t^n}{n!},\qquad N(D)=(4D^3+D+3)^2. \end{aligned}(8) Write ι(t)\iota(t) for the image of a nonnegative real number in [0,∞][0,\infty].

Theorem 1. Let D≥2D\ge2 be a natural number, and suppose every vertex of Γ\Gamma has at most DD neighbors. Let QQ be an equivariant marked subdivision pair, let μ\mu be a measure on XX, and let B⊆XB\subseteq X. If the probability center cc has support of cardinality at most n+1n+1, then, for every real b≥0b\ge0, the top-degree ball mass satisfies (9): MnB(c,b)≤(n+1)D⌊bn⌋++1D2nι(1n!)μ(B).M_n^B(c,b)\le (n+1)D^{\lfloor b\sqrt n\rfloor_++1}D^{2n} \iota\!\left(\frac1{n!}\right)\mu(B).(9)✦ For natural numbers D≥2D\ge2 and n≥N(D)n\ge N(D), the numerical estimate is (10): (n+1)D⌊4D2n⌋++1D2nn!<(2D2)nn!.\frac{(n+1)D^{\lfloor4D^2\sqrt n\rfloor_++1}D^{2n}}{n!} <\frac{(2D^2)^n}{n!}.(10)✦ For natural numbers D≥2D\ge2 and n≥N(D)n\ge N(D), the ball volume bound satisfies (11): VD,n(4D2)<(2D2)[n].V_{D,n}(4D^2)<(2D^2)^{[n]}.(11)✦ Put R=4D2R=4D^2 and r=2D2r=2D^2.

In the paper, the result reads: for degree at most D≥2D\ge2, R=4D2R=4D^2 and r=R/2r=R/2, the total nn-area of LR(c)L_R(c) in top interiors is at most VR=(n+1)D⌊Rn⌋+1D2n/n!V_R=(n+1)D^{\lfloor R\sqrt n\rfloor+1}D^{2n}/n!, and VR<(R−r)n/n!V_R<(R-r)^n/n! for all sufficiently large nn depending only on DD.

What Is Checked

The ball mass estimate is stated over a connected graph of degree at most a natural number D≥2D\ge2, its simplicial realization, an equivariant marked subdivision pair, a measure on the parameter space and an arbitrary parameter subset, a finitely supported probability center with at most n+1n+1 support vertices, and a nonnegative radius. The strict numerical estimates are stated for natural numbers D≥2D\ge2 and n≥N(D)n\ge N(D).