OctoInformBeta

OpenAI preprint, September 2026 · §5. Contractions into ℓ₁

Smoothing, interface jumps, and gradient budgets: Bounds via interface jumps and good-chart averaging for product mollifiers κλ\kappa_\lambda with 0<λ<10 < \lambda < 1, with contraction weight and constants

Mollified chart step functions satisfy error and gradient 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

For natural numbers m,S,Nm, S, N, suppose SS charts on Euclidean space E=ℝm+1E = \mathbb{R}^{m+1} with Euclidean norm ∥⋅∥\|\cdot\|, each cut by NN families of parallel hyperplanes at spacing τ>0\tau > 0, carry a map FF into ℝA\mathbb{R}^A for a finite coordinate set AA. Extending FF to piecewise-constant chart step functions fs,af_{s,a} on EE for 1≤s≤S1 \le s \le S and a∈Aa \in A creates jump discontinuities across the threshold hyperplanes, so convolving each component with a product mollifier κλ\kappa_\lambda of width λ∈(0,1)\lambda \in (0, 1) yields smooth functions Hs,aH_{s,a} on EE. The central question is how closely Hs,aH_{s,a} approximates fs,af_{s,a} and how the total gradient ∑a∈A∥∇Hs,a(θ)∥\sum_{a \in A} \|\nabla H_{s,a}(\theta)\| at θ∈E\theta \in E is controlled by cell-interface jumps and cross-chart mismatches when direction vectors are small against dual vectors on at least half of the charts.

Chart Step Functions

Equip E=ℝm+1E = \mathbb{R}^{m+1} with the standard inner product ⟨⋅,⋅⟩\langle \cdot, \cdot \rangle, writing coordinates as θ=(θ0,θ1,…,θm)\theta = (\theta_0, \theta_1, \dots, \theta_m). For t∈ℝt \in \mathbb{R}, the cube of radius tt in EE is Qt={θ∈E:|θj|<t for all 0≤j≤m}Q_t = \{\theta \in E : |\theta_j| < t \text{ for all } 0 \le j \le m\}.✦ Given a family u=(u1,…,uN)u = (u_1, \dots, u_N) of vectors in EE and a scale τ∈ℝ\tau \in \mathbb{R}, a point θ∈E\theta \in E is τ\tau-regular for uu if ⟨ui,θ⟩≠kτ\langle u_i, \theta \rangle \neq k\tau for all 1≤i≤N1 \le i \le N and k∈ℤk \in \mathbb{Z}✦, and its integer label is X(u,τ,θ)∈ℤNX(u, \tau, \theta) \in \mathbb{Z}^N with entries X(u,τ,θ)i=⌊⟨ui,θ⟩/τ⌋X(u, \tau, \theta)_i = \lfloor \langle u_i, \theta \rangle / \tau \rfloor.✦ Given SS families u=(us,i)1≤s≤S,1≤i≤Nu = (u_{s,i})_{1 \le s \le S, \, 1 \le i \le N} in EE, write us=(us,1,…,us,N)u_s = (u_{s,1}, \dots, u_{s,N}) and Xs(θ)=X(us,τ,θ)X_s(\theta) = X(u_s, \tau, \theta); the rounded chart vertex set V(u,τ)V(u, \tau) consists of all pairs (s,x)(s, x) with 1≤s≤S1 \le s \le S and x∈ℤNx \in \mathbb{Z}^N for which some τ\tau-regular point θ∈Q2\theta \in Q_2 for usu_s satisfies Xs(θ)=xX_s(\theta) = x✦, and each such θ\theta defines a chart vertex vs(θ)=(s,Xs(θ))∈V(u,τ)v_s(\theta) = (s, X_s(\theta)) \in V(u, \tau).✦

Let Z=ℝmZ = \mathbb{R}^m and identify EE with ℝ×Z\mathbb{R} \times Z via the linear homeomorphism σ(θ)=(θ0,θ′)\sigma(\theta) = (\theta_0, \theta'), where θ′=(θ1,…,θm)∈Z\theta' = (\theta_1, \dots, \theta_m) \in Z.✦ For any vector w∈Ew \in E, the non-pivot linear functional rw:Z→ℝr_w : Z \to \mathbb{R} on ZZ is defined by rw(z)=⟨w,σ−1(0,z)⟩r_w(z) = \langle w, \sigma^{-1}(0, z) \rangle✦, and when w0≠0w_0 \neq 0, the hyperplane slope functional bw:Z→ℝb_w : Z \to \mathbb{R} on ZZ is defined by bw(z)=−w0−1rw(z)b_w(z) = -w_0^{-1} r_w(z).✦ For a finite index set II and vectors u=(ui)i∈Iu = (u_i)_{i \in I} in EE with nonzero pivot coordinates pi=ui,0p_i = u_{i,0}, the slope family (Bi)i∈I(B_i)_{i \in I} of uu is the family of linear functionals on ZZ defined by Bi=buiB_i = b_{u_i} for each i∈Ii \in I.✦ On ℝ×Z\mathbb{R} \times Z, each slope functional B′:Z→ℝB' : Z \to \mathbb{R} defines the slice linear form ϕB′:ℝ×Z→ℝ\phi_{B'} : \mathbb{R} \times Z \to \mathbb{R} by ϕB′(t,z)=t−B′(z)\phi_{B'}(t, z) = t - B'(z)✦, and the pivoted slice functionals ℓi:ℝ×Z→ℝ\ell_i : \mathbb{R} \times Z \to \mathbb{R} are defined by ℓi(t,z)=piϕBi(t,z)\ell_i(t, z) = p_i \phi_{B_i}(t, z), which satisfy ℓi(σ(θ))=⟨ui,θ⟩\ell_i(\sigma(\theta)) = \langle u_i, \theta \rangle.✦

The split cube domain Ω⊆ℝ×Z\Omega \subseteq \mathbb{R} \times Z is the image subset Ω=σ(Q2)\Omega = \sigma(Q_2).✦ For a family of continuous linear functionals ℓ=(ℓi)i∈I\ell = (\ell_i)_{i \in I} on ℝ×Z\mathbb{R} \times Z and a subset 𝒰⊆ℝ×Z\mathcal{U} \subseteq \mathbb{R} \times Z, a point y∈ℝ×Zy \in \mathbb{R} \times Z is τ\tau-regular if ℓi(y)≠kτ\ell_i(y) \neq k\tau for all i∈Ii \in I and k∈ℤk \in \mathbb{Z}✦, with integer label (⌊ℓi(y)/τ⌋)i∈I(\lfloor \ell_i(y)/\tau \rfloor)_{i \in I}✦; the label vertex set Λ(ℓ,τ,𝒰)\Lambda(\ell, \tau, \mathcal{U}) consists of all integer labels realized by τ\tau-regular points in 𝒰\mathcal{U}, each associated with its label vertex.✦ For the pivoted slice functionals ℓ=(ℓs,i)1≤i≤N\ell = (\ell_{s,i})_{1 \le i \le N} of usu_s with us,i,0≠0u_{s,i,0} \neq 0, σ\sigma maps Q2Q_2 onto Ω\Omega and identifies Λ(ℓ,τ,Ω)\Lambda(\ell, \tau, \Omega) with the chart-ss vertices of V(u,τ)V(u, \tau).

Given a finite threshold set 𝒦⊆ℤ\mathcal{K} \subseteq \mathbb{Z}, a scale τ>0\tau > 0, continuous linear functionals ℓ=(ℓi)i∈I\ell = (\ell_i)_{i \in I} on ℝ×Z\mathbb{R} \times Z, and nonzero pivots (pi)i∈I(p_i)_{i \in I}, the sign pattern of a point y∈ℝ×Zy \in \mathbb{R} \times Z is the subset of I×𝒦I \times \mathcal{K} defined by {(i,k)∈I×𝒦:(ℓi(y)−kτ)/pi<0}\{(i, k) \in I \times \mathcal{K} : (\ell_i(y) - k\tau)/p_i < 0\}.✦ A pattern extension assigns to each map F0:Λ(ℓ,τ,𝒰)→ℝAF_0 : \Lambda(\ell, \tau, \mathcal{U}) \to \mathbb{R}^A a map F̃0\widetilde{F}_0 defined on all subsets P⊆I×𝒦P \subseteq I \times \mathcal{K} and valued in ℝA\mathbb{R}^A✦; specializing to the pivoted slice functionals and pivots of uu gives the cell pattern values.✦ Given slopes (Bi)i∈I(B_i)_{i \in I}, offsets (qi,k)i∈I,k∈𝒦(q_{i,k})_{i \in I, k \in \mathcal{K}}, and a real-valued function Φ\Phi on subsets P⊆I×𝒦P \subseteq I \times \mathcal{K}, the slice cell function sends (t,z)∈ℝ×Z(t, z) \in \mathbb{R} \times Z to Φ({(i,k)∈I×𝒦:ϕBi(t,z)+qi,k<0})\Phi(\{(i,k) \in I \times \mathcal{K} : \phi_{B_i}(t, z) + q_{i,k} < 0\})✦, and setting qi,k=−kτ/piq_{i,k} = -k\tau / p_i and composing with σ\sigma yields the Euclidean chart function for uu, τ\tau, 𝒦\mathcal{K}, and 𝒰\mathcal{U}.✦ The chart label range 𝒦τ⊆ℤ\mathcal{K}_\tau \subseteq \mathbb{Z} is defined by (1): 𝒦τ={k∈ℤ:−⌈4m+1τ+1⌉≤k≤⌈4m+1τ+1⌉},\mathcal{K}_\tau = \left\{ k \in \mathbb{Z} : -\left\lceil \frac{4\sqrt{m+1}}{\tau} + 1 \right\rceil \le k \le \left\lceil \frac{4\sqrt{m+1}}{\tau} + 1 \right\rceil \right\},(1)✦ For any collection u=(us,i)1≤s≤S,1≤i≤Nu = (u_{s,i})_{1 \le s \le S, \, 1 \le i \le N} with us,i,0≠0u_{s,i,0} \neq 0 and vertex map F:V(u,τ)→ℝAF : V(u, \tau) \to \mathbb{R}^A, let Bs,i=bus,iB_{s,i} = b_{u_{s,i}} and let F̃s\widetilde{F}_s be the cell pattern extension of x↦F(s,x)x \mapsto F(s, x) from the chart-ss slice label vertices over Ω\Omega to subsets of {1,…,N}×𝒦τ\{1, \dots, N\} \times \mathcal{K}_\tau; then the concrete chart step function fs,a:E→ℝf_{s,a} : E \to \mathbb{R} for 1≤s≤S1 \le s \le S and a∈Aa \in A is defined by (2): fs,a(θ)=F̃s({(i,k)∈{1,…,N}×𝒦τ:θ0−Bs,i(θ′)−kτus,i,0<0})a,f_{s,a}(\theta) = \widetilde{F}_s\!\left(\left\{(i, k) \in \{1, \dots, N\} \times \mathcal{K}_\tau : \theta_0 - B_{s,i}(\theta') - \frac{k\tau}{u_{s,i,0}} < 0\right\}\right)_a,(2)✦

Mollifiers and Constants

Let ψ:ℝ→ℝ\psi : \mathbb{R} \to \mathbb{R} be the smooth (C∞C^\infty) bump function centered at 00 with inner radius 1/21/2 and outer radius 11, equal to 11 on [−1/2,1/2][-1/2, 1/2] and supported in (−1,1)(-1, 1).✦ Normalizing ψ\psi to unit Lebesgue integral produces the standard mollifier ρ:ℝ→ℝ\rho : \mathbb{R} \to \mathbb{R}, which is smooth, nonnegative, supported in (−1,1)(-1, 1), and of integral 11.✦

For a width λ>0\lambda > 0, the scaled mollifier density ρλ:ℝ→ℝ\rho_\lambda : \mathbb{R} \to \mathbb{R} is the function defined by ρλ(t)=λ−1ρ(t/λ)\rho_\lambda(t) = \lambda^{-1}\rho(t/\lambda).✦ Given any φ:ℝ→ℝ\varphi : \mathbb{R} \to \mathbb{R}, the product kernel Kφ:E→ℝK_\varphi : E \to \mathbb{R} on EE is defined by Kφ(θ)=∏j=0mφ(θj)K_\varphi(\theta) = \prod_{j=0}^m \varphi(\theta_j)✦, and taking φ=ρλ\varphi = \rho_\lambda yields the product mollifier kernel κλ:E→ℝ\kappa_\lambda : E \to \mathbb{R} in (3): κλ(θ)=∏j=0mλ−1ρ(θjλ).\kappa_\lambda(\theta) = \prod_{j=0}^m \lambda^{-1}\rho\!\left(\frac{\theta_j}{\lambda}\right).(3)✦ For λ>0\lambda > 0, the mollifier Lipschitz constant L(λ)∈ℝL(\lambda) \in \mathbb{R} is given by (4): L(λ)=λ−1∫ℝ|ρ′(t)|dt.L(\lambda) = \lambda^{-1} \int_{\mathbb{R}} |\rho'(t)|\,dt.(4)✦

For a kernel κ:E→ℝ\kappa : E \to \mathbb{R} and a function f:E→ℝf : E \to \mathbb{R}, the convolution smoothing κ*f:E→ℝ\kappa * f : E \to \mathbb{R} with respect to Lebesgue measure on EE is defined by (κ*f)(θ)=∫Eκ(η)f(θ−η)dη(\kappa * f)(\theta) = \int_E \kappa(\eta) f(\theta - \eta)\,d\eta.✦ For each chart 1≤s≤S1 \le s \le S and coordinate a∈Aa \in A, the smoothed chart function Hs,a:E→ℝH_{s,a} : E \to \mathbb{R} is the convolution Hs,a=κλ*fs,aH_{s,a} = \kappa_\lambda * f_{s,a} defined by (5): Hs,a(θ)=(κλ*fs,a)(θ)=∫Eκλ(η)fs,a(θ−η)dη.H_{s,a}(\theta) = (\kappa_\lambda * f_{s,a})(\theta) = \int_E \kappa_\lambda(\eta) f_{s,a}(\theta - \eta)\,d\eta.(5)

The profile function g:ℝ→ℝg : \mathbb{R} \to \mathbb{R} is given by g(t)=1−exp⁡(−1/(2t2))g(t) = 1 - \exp(-1/(2t^2))✦, the profile constant is the real number c*=2∫0∞g(t)dtc^* = 2 \int_0^\infty g(t)\,dt✦, and the source metric constant is the real number C0=9+2c*C_0 = 9 + 2c^*.✦

The Claim

Definition 1. The mollifier derivative integral H∈ℝH \in \mathbb{R} is defined by (6): H=∫ℝ|ρ′(t)|dt.H = \int_{\mathbb{R}} |\rho'(t)|\,dt.(6)✦ In terms of c*c^*, C0C_0, and HH, the contraction constant C∈ℝC \in \mathbb{R} is given by (7): C=2(c*+C0)+π(1100000000+C0H)+1.C = 2(c^* + C_0) + \pi(1100000000 + C_0 H) + 1.(7)✦ The contraction weight function W:ℕ→ℝW : \mathbb{N} \to \mathbb{R} maps each natural number mm to the real number (8): W(m)=m(log⁡m)2log⁡m.W(m) = \sqrt{m}\,(\log m)^2 \sqrt{\log m}.(8)✦

Theorem 1. Let m,S,N∈ℕm, S, N \in \mathbb{N} and let AA be a finite set. Suppose vectors u=(us,i)1≤s≤S,1≤i≤Nu = (u_{s,i})_{1 \le s \le S, \, 1 \le i \le N} in E=ℝm+1E = \mathbb{R}^{m+1} satisfy us,i,0≠0u_{s,i,0} \neq 0 and ∥us,i∥≤2\|u_{s,i}\| \le 2 for all ss and ii, let τ>0\tau > 0 and 0<λ<10 < \lambda < 1, and let F:V(u,τ)→ℝAF : V(u, \tau) \to \mathbb{R}^A satisfy |F(v)a|≤M|F(v)_a| \le M for some M≥0M \ge 0 and all v∈V(u,τ)v \in V(u, \tau), a∈Aa \in A. Suppose further that there is R∈ℝR \in \mathbb{R} such that ∑a∈A|F(vs(θ))a−F(vt(θ))a|≤R\sum_{a \in A} |F(v_s(\theta))_a - F(v_t(\theta))_a| \le R whenever 1≤s,t≤S1 \le s, t \le S and θ∈Q2\theta \in Q_2 is τ\tau-regular for both usu_s and utu_t. Then for all 1≤s,t≤S1 \le s, t \le S and every θ∈[−1,1]m+1\theta \in [-1, 1]^{m+1}, the gradients of the smoothed chart functions satisfy (9): ∑a∈A∥∇Hs,a(θ)−∇Ht,a(θ)∥≤(m+1)RL(λ).\sum_{a \in A} \|\nabla H_{s,a}(\theta) - \nabla H_{t,a}(\theta)\| \le (m + 1) R L(\lambda).(9)✦

Theorem 2. Given m,S,N∈ℕm, S, N \in \mathbb{N}, a finite set AA, vectors u=(us,i)1≤s≤S,1≤i≤Nu = (u_{s,i})_{1 \le s \le S, \, 1 \le i \le N} in E=ℝm+1E = \mathbb{R}^{m+1} with us,i,0≠0u_{s,i,0} \neq 0 and ∥us,i∥≤2\|u_{s,i}\| \le 2, scales τ>0\tau > 0 and 0<λ<10 < \lambda < 1, and a map F:V(u,τ)→ℝAF : V(u, \tau) \to \mathbb{R}^A satisfying |F(v)a|≤M|F(v)_a| \le M for some M≥0M \ge 0, suppose G≥0G \ge 0 and R∈ℝR \in \mathbb{R} are such that ∑a∈A|F(vs(θ))a−F(vs(η))a|≤G∥θ−η∥+R\sum_{a \in A} |F(v_s(\theta))_a - F(v_s(\eta))_a| \le G \|\theta - \eta\| + R for every 1≤s≤S1 \le s \le S and all θ,η∈Q2\theta, \eta \in Q_2 that are τ\tau-regular for usu_s. Then for any 1≤s≤S1 \le s \le S and any θ∈[−1,1]m+1\theta \in [-1, 1]^{m+1} that is τ\tau-regular for usu_s, the smoothing error satisfies (10): ∑a∈A|Hs,a(θ)−fs,a(θ)|≤Gλm+1+R.\sum_{a \in A} |H_{s,a}(\theta) - f_{s,a}(\theta)| \le G \lambda \sqrt{m + 1} + R.(10)✦

Theorem 3. Take m,S,N∈ℕm, S, N \in \mathbb{N}, a finite set AA, scales τ>0\tau > 0 and 0<λ<10 < \lambda < 1, parameters δ≥0\delta \ge 0, J≥0J \ge 0, and R∈ℝR \in \mathbb{R}, and vectors u=(us,i)1≤s≤S,1≤i≤Nu = (u_{s,i})_{1 \le s \le S, \, 1 \le i \le N} in E=ℝm+1E = \mathbb{R}^{m+1} satisfying us,i,0≠0u_{s,i,0} \neq 0, ∥us,i∥≤2\|u_{s,i}\| \le 2, and us,i≠tus,ju_{s,i} \neq t u_{s,j} whenever i≠ji \neq j and t∈ℝt \in \mathbb{R}. Let F:V(u,τ)→ℝAF : V(u, \tau) \to \mathbb{R}^A satisfy |F(v)a|≤M|F(v)_a| \le M for some M≥0M \ge 0 and obey the cross-chart bound ∑a∈A|F(vs(θ))a−F(vt(θ))a|≤R\sum_{a \in A} |F(v_s(\theta))_a - F(v_t(\theta))_a| \le R for all s,ts, t and θ∈Q2\theta \in Q_2 that is τ\tau-regular for usu_s and utu_t, together with the interface jump bound ∑a∈A|F(v+)a−F(v−)a|≤J\sum_{a \in A} |F(v_+)_a - F(v_-)_a| \le J whenever v+=(s,x+)v_+ = (s, x_+) and v−=(s,x−)v_- = (s, x_-) in V(u,τ)V(u, \tau) share a chart ss and satisfy (x+)i=(x−)i+1(x_+)_i = (x_-)_i + 1 at one index ii and (x+)j=(x−)j(x_+)_j = (x_-)_j for all j≠ij \neq i. Assume also that τ|us,i,0|L(λ)≤1\frac{\tau}{|u_{s,i,0}|} L(\lambda) \le 1 for all ss and ii, and that for every continuous linear functional DD on EE there exists a subset 𝒢⊆{1,…,S}\mathcal{G} \subseteq \{1, \dots, S\} with |𝒢|≥S/2|\mathcal{G}| \ge S/2 such that |D(us,i)|≤δ∥D∥|D(u_{s,i})| \le \delta \|D\| for all s∈𝒢s \in \mathcal{G} and ii. Then for every chart 1≤s0≤S1 \le s_0 \le S and every θ∈[−1,1]m+1\theta \in [-1, 1]^{m+1}, the smoothed gradient satisfies (11): ∑a∈A∥∇Hs0,a(θ)∥≤2(δN2Jτ+(m+1)RL(λ)).\sum_{a \in A} \|\nabla H_{s_0, a}(\theta)\| \le 2 \left( \delta N \frac{2J}{\tau} + (m + 1) R L(\lambda) \right).(11)✦

Theorem 4. In any real inner-product space ℰ\mathcal{E}, let 𝒮\mathcal{S} and II be finite index sets, let us,i,a,es∈ℰu_{s,i}, a, e_s \in \mathcal{E} and βs,i∈ℝ\beta_{s,i} \in \mathbb{R} for s∈𝒮s \in \mathcal{S} and i∈Ii \in I, and let δ≥0\delta \ge 0. Suppose a+es=∑i∈Iβs,ius,ia + e_s = \sum_{i \in I} \beta_{s,i} u_{s,i} for every s∈𝒮s \in \mathcal{S}, and suppose a subset 𝒢⊆𝒮\mathcal{G} \subseteq \mathcal{S} with |𝒢|≥|𝒮|/2|\mathcal{G}| \ge |\mathcal{S}|/2 satisfies |⟨us,i,a⟩|≤δ∥a∥|\langle u_{s,i}, a \rangle| \le \delta \|a\| for all s∈𝒢s \in \mathcal{G} and i∈Ii \in I. Then (12) holds: |𝒮|∥a∥≤2∑s∈𝒮(δ∑i∈I|βs,i|+∥es∥).|\mathcal{S}| \|a\| \le 2 \sum_{s \in \mathcal{S}} \left( \delta \sum_{i \in I} |\beta_{s,i}| + \|e_s\| \right).(12)✦

Theorem 5. Suppose 𝒮\mathcal{S}, II, and AA are finite index sets with |𝒮|>0|\mathcal{S}| > 0, ℰ\mathcal{E} is a real inner-product space, and δ≥0\delta \ge 0, B,R∈ℝB, R \in \mathbb{R}. Let us,i,aα,es,α∈ℰu_{s,i}, a_\alpha, e_{s,\alpha} \in \mathcal{E} and βs,i,α∈ℝ\beta_{s,i,\alpha} \in \mathbb{R} for s∈𝒮s \in \mathcal{S}, i∈Ii \in I, and α∈A\alpha \in A satisfy aα+es,α=∑i∈Iβs,i,αus,ia_\alpha + e_{s,\alpha} = \sum_{i \in I} \beta_{s,i,\alpha} u_{s,i} for all s,αs, \alpha. Assume that for every α∈A\alpha \in A there exists a subset 𝒢α⊆𝒮\mathcal{G}_\alpha \subseteq \mathcal{S} with |𝒢α|≥|𝒮|/2|\mathcal{G}_\alpha| \ge |\mathcal{S}|/2 such that |⟨us,i,aα⟩|≤δ∥aα∥|\langle u_{s,i}, a_\alpha \rangle| \le \delta \|a_\alpha\| for all s∈𝒢αs \in \mathcal{G}_\alpha and i∈Ii \in I, and that for every s∈𝒮s \in \mathcal{S} the budgets ∑i∈I∑α∈A|βs,i,α|≤B\sum_{i \in I} \sum_{\alpha \in A} |\beta_{s,i,\alpha}| \le B and ∑α∈A∥es,α∥≤R\sum_{\alpha \in A} \|e_{s,\alpha}\| \le R hold. Then the total norm satisfies (13): ∑α∈A∥aα∥≤2(δB+R).\sum_{\alpha \in A} \|a_\alpha\| \le 2(\delta B + R).(13)✦

What Is Checked

The definitions (6), (7), and (8) and the bounds (9), (10), and (11) are stated over the standard mollifier ρ\rho, the constants c*c^* and C0C_0, and the smoothed chart step functions Hs,a=κλ*fs,aH_{s,a} = \kappa_\lambda * f_{s,a} on E=ℝm+1E = \mathbb{R}^{m+1} for bounded vertex maps FF and direction families with nonzero zeroth coordinates and ∥us,i∥≤2\|u_{s,i}\| \le 2 at scales τ>0\tau > 0 and 0<λ<10 < \lambda < 1, whereas (12) and (13) are stated over finite vector decompositions in an arbitrary real inner-product space.