OctoInformBeta

OpenAI preprint, September 2026 · §3. A Hilbert kernel with chart extensions

Lemma 3.2: Linearity alongside residual and ℓ1\ell^1 inequalities for chart extensions, and variation control from pairing differentials

Chart extension bounds and Hilbert kernel segment estimates.

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

Consider nonnegative integers SS and NN together with SS families of NN vectors in a real inner-product space EE with inner product ⟨⋅,⋅⟩\langle \cdot, \cdot \rangle and norm ∥⋅∥\|\cdot\|, and take NN-dimensional Euclidean space F=ℝNF = \mathbb{R}^N equipped with the Euclidean norm ∥⋅∥\|\cdot\| as the coordinate target; in the uniform sparsest-cut construction, S=m3S = m^3 and N=m6N = m^6 for a large positive integer mm. When EE is paired with FF via frame chart operators and continuous linear maps from FF to EE, the questions are what linearity and residual or ℓ1\ell^1 norm bounds govern the resulting chart extension, and how an estimate on the kernel pairing differential across a scale interval bounds the change in the Hilbert kernel along a line segment.

Chart Operators and Extensions

For a nonnegative integer NN with finite index set [N][N] of size NN, a real inner-product space EE, and a family u=(ui)i∈[N]u = (u_i)_{i \in [N]} of vectors in EE, the frame chart operator Ju:E→FJ_u : E \to F into NN-dimensional Euclidean space F=ℝNF = \mathbb{R}^N is the continuous linear map defined by (1): (Juθ)i=⟨ui,θ⟩(θ∈E,i∈[N]).(J_u \theta)_i = \langle u_i, \theta \rangle \qquad (\theta \in E,\ i \in [N]).(1)✦

For continuous linear maps J:E→FJ : E \to F and B:F→EB : F \to E between real normed spaces EE and FF, and the identity operator II on FF, the residual operator I−JB:F→FI - J B : F \to F is the continuous linear operator on FF given by (2): (I−JB)x=x−J(Bx)(x∈F).(I - J B) x = x - J(B x) \qquad (x \in F).(2)✦

The chart extension αg,B,q,k,ε\alpha_{g,B,q,k,\varepsilon} associated with nonnegative integers SS and NN with index sets [S][S] and [N][N], SS families g=(gs,i)s∈[S],i∈[N]g = (g_{s,i})_{s \in [S],\, i \in [N]} of NN vectors in a real inner-product space EE, continuous linear maps B=(Bs)s∈[S]B = (B_s)_{s \in [S]} from F=ℝNF = \mathbb{R}^N to EE, and real scalars q,k,ε∈ℝq, k, \varepsilon \in \mathbb{R} is the real-valued function of a chart index s∈[S]s \in [S], a vector x∈Fx \in F, and an index pair j=(j1,j2)∈[S]×[N]j = (j_1, j_2) \in [S] \times [N] given by (3): αg,B,q,k,ε(s,x)j={kε(xj2−⟨gj1,j2,Bsx⟩q)if j1=s,0if j1≠s.\alpha_{g,B,q,k,\varepsilon}(s, x)_j = \begin{cases} \dfrac{k}{\varepsilon}\left(x_{j_2} - \dfrac{\langle g_{j_1, j_2}, B_s x \rangle}{q}\right) & \text{if } j_1 = s, \\[8pt] 0 & \text{if } j_1 \ne s. \end{cases}(3)✦

The Hilbert Kernel and Its Differential

Given a finite index set ι\iota, a vector family g=(gi)i∈ιg = (g_i)_{i \in \iota} in a real inner-product space EE, real scalars q,ε,σ∈ℝq, \varepsilon, \sigma \in \mathbb{R}, points θ,η∈E\theta, \eta \in E, and coefficient functions α,β:ι→ℝ\alpha, \beta : \iota \to \mathbb{R}, the kernel pairing Φg,q,ε(σ;θ,η,α,β)\Phi_{g,q,\varepsilon}(\sigma; \theta, \eta, \alpha, \beta) is the real scalar defined by (4): Φg,q,ε(σ;θ,η,α,β)=qexp⁡(−∥η−θ∥22σ2)+εσ∑i∈ι((1+αiβi)cos⁡⟨gi,η−θ⟩σ+(αi−βi)sin⁡⟨gi,η−θ⟩σ).\begin{aligned} \Phi_{g,q,\varepsilon}(\sigma; \theta, \eta, \alpha, \beta) &= q \exp\!\left(-\frac{\|\eta - \theta\|^2}{2\sigma^2}\right) \\ &\quad + \frac{\varepsilon}{\sigma} \sum_{i \in \iota} \biggl((1 + \alpha_i \beta_i)\cos\frac{\langle g_i, \eta - \theta \rangle}{\sigma} + (\alpha_i - \beta_i)\sin\frac{\langle g_i, \eta - \theta \rangle}{\sigma}\biggr). \end{aligned}(4)✦

At scale endpoints a,b∈ℝa, b \in \mathbb{R}, the Hilbert kernel Kg,q,ε,a,b(θ,η,α,β)K_{g,q,\varepsilon,a,b}(\theta, \eta, \alpha, \beta) associated with a finite index set ι\iota, a vector family g=(gi)i∈ιg = (g_i)_{i \in \iota} in a real inner-product space EE, real weights q,ε∈ℝq, \varepsilon \in \mathbb{R}, points θ,η∈E\theta, \eta \in E, and coefficient functions α,β:ι→ℝ\alpha, \beta : \iota \to \mathbb{R} is the real scalar defined by (5): Kg,q,ε,a,b(θ,η,α,β)=∫abΦg,q,ε(σ;θ,η,α,β)dσ.K_{g,q,\varepsilon,a,b}(\theta, \eta, \alpha, \beta) = \int_a^b \Phi_{g,q,\varepsilon}(\sigma; \theta, \eta, \alpha, \beta)\,d\sigma.(5)✦

The kernel pairing differential Ψg,q,ε(σ;θ,η,v,α,β,δ)\Psi_{g,q,\varepsilon}(\sigma; \theta, \eta, v, \alpha, \beta, \delta) associated with a finite index set ι\iota, a vector family g=(gi)i∈ιg = (g_i)_{i \in \iota} in a real inner-product space EE, real scalars q,ε,σ∈ℝq, \varepsilon, \sigma \in \mathbb{R}, points θ,η,v∈E\theta, \eta, v \in E, and coefficient functions α,β,δ:ι→ℝ\alpha, \beta, \delta : \iota \to \mathbb{R} is the real scalar given by (6): Ψg,q,ε(σ;θ,η,v,α,β,δ)=qσ2⟨η−θ,v⟩exp⁡(−∥η−θ∥22σ2)+εσ∑i∈ι(δi(βicos⁡⟨gi,η−θ⟩σ+sin⁡⟨gi,η−θ⟩σ)+⟨gi,v⟩σ((1+αiβi)sin⁡⟨gi,η−θ⟩σ−(αi−βi)cos⁡⟨gi,η−θ⟩σ)).\begin{aligned} \Psi_{g,q,\varepsilon}(\sigma; \theta, \eta, v, \alpha, \beta, \delta) &= \frac{q}{\sigma^2}\langle \eta - \theta, v \rangle \exp\!\left(-\frac{\|\eta - \theta\|^2}{2\sigma^2}\right) \\ &\quad + \frac{\varepsilon}{\sigma}\sum_{i \in \iota} \biggl( \delta_i \Bigl(\beta_i \cos\frac{\langle g_i, \eta - \theta \rangle}{\sigma} + \sin\frac{\langle g_i, \eta - \theta \rangle}{\sigma}\Bigr) \\ &\qquad + \frac{\langle g_i, v \rangle}{\sigma} \Bigl((1 + \alpha_i \beta_i)\sin\frac{\langle g_i, \eta - \theta \rangle}{\sigma} - (\alpha_i - \beta_i)\cos\frac{\langle g_i, \eta - \theta \rangle}{\sigma}\Bigr) \biggr). \end{aligned}(6)

The Claim

Theorem 1. Let SS and NN be nonnegative integers with index sets [S][S] and [N][N], let g=(gs,i)s∈[S],i∈[N]g = (g_{s,i})_{s \in [S],\, i \in [N]} be families of vectors in a real inner-product space EE, let B=(Bs)s∈[S]B = (B_s)_{s \in [S]} be continuous linear maps from NN-dimensional Euclidean space F=ℝNF = \mathbb{R}^N with standard basis (ei)i∈[N](e_i)_{i \in [N]} to EE, let q,k,ε∈ℝq, k, \varepsilon \in \mathbb{R}, and fix s∈[S]s \in [S] and j∈[S]×[N]j \in [S] \times [N]. For all x,h∈Fx, h \in F and t∈ℝt \in \mathbb{R}, the chart extension satisfies (7): αg,B,q,k,ε(s,x+th)j=αg,B,q,k,ε(s,x)j+tαg,B,q,k,ε(s,h)j.\alpha_{g,B,q,k,\varepsilon}(s, x + t h)_j = \alpha_{g,B,q,k,\varepsilon}(s, x)_j + t\,\alpha_{g,B,q,k,\varepsilon}(s, h)_j.(7)✦ Writing q−1gsq^{-1} g_s for the family i↦q−1gs,ii \mapsto q^{-1} g_{s,i}, if k≥0k \ge 0, ε>0\varepsilon > 0, and x∈Fx \in F satisfies ∥(I−Jq−1gsBs)x∥≤R\|(I - J_{q^{-1} g_s} B_s) x\| \le R for some R∈ℝR \in \mathbb{R}, then the norm bound (8) holds: |αg,B,q,k,ε(s,x)j|≤kεR.|\alpha_{g,B,q,k,\varepsilon}(s, x)_j| \le \frac{k}{\varepsilon} R.(8)✦ If q>0q > 0, 0≤k≤10 \le k \le 1, ε>0\varepsilon > 0, h=(hi)i∈[N]∈Fh = (h_i)_{i \in [N]} \in F satisfies ∑i∈[N]|hi|≤H\sum_{i \in [N]} |h_i| \le H for some H∈ℝH \in \mathbb{R}, and ∥gs,i∥≤2q\|g_{s,i}\| \le 2q and ∥Bsei∥≤4k\|B_s e_i\| \le 4k for every i∈[N]i \in [N], then the component bound (9) holds: |αg,B,q,k,ε(s,h)j|≤9kεH.|\alpha_{g,B,q,k,\varepsilon}(s, h)_j| \le \frac{9k}{\varepsilon} H.(9)✦

Let ι\iota be a finite index set, let g=(gi)i∈ιg = (g_i)_{i \in \iota} be vectors in a real inner-product space EE, let q,ε,a,b,C,D∈ℝq, \varepsilon, a, b, C, D \in \mathbb{R} satisfy 0<a≤b0 < a \le b and D≥0D \ge 0, let θ,η,v∈E\theta, \eta, v \in E, and let α,β,δ:ι→ℝ\alpha, \beta, \delta : \iota \to \mathbb{R}. If |Ψg,q,ε(σ;θ+tv,η,v,α+tδ,β,δ)|≤C/σ+D/σ2|\Psi_{g,q,\varepsilon}(\sigma; \theta + t v, \eta, v, \alpha + t\delta, \beta, \delta)| \le C/\sigma + D/\sigma^2 for all t∈[0,1]t \in [0, 1] and σ∈[a,b]\sigma \in [a, b], then the segment bound (10) holds: |Kg,q,ε,a,b(θ+v,η,α+δ,β)−Kg,q,ε,a,b(θ,η,α,β)|≤Cln⁡ba+Da.|K_{g,q,\varepsilon,a,b}(\theta + v, \eta, \alpha + \delta, \beta) - K_{g,q,\varepsilon,a,b}(\theta, \eta, \alpha, \beta)| \le C \ln\frac{b}{a} + \frac{D}{a}.(10)✦

In the paper, the result reads: for a map p:ℝm→ℋp : \mathbb{R}^m \to \mathcal{H} into a Hilbert space ℋ\mathcal{H}, a constant CC, the operator norm ∥⋅∥op\|\cdot\|_{\mathrm{op}}, and the frame chart operator JsJ_s of the normalized directions us,i=gs,i/mu_{s,i} = g_{s,i}/\sqrt{m} with gs,i∈ℝmg_{s,i} \in \mathbb{R}^m, S=m3S = m^3, and N=m6N = m^6, every 1≤s≤S1 \le s \le S admits a smooth map As:ℝm→ℒ(ℝN,ℋ)A_s : \mathbb{R}^m \to \mathcal{L}(\mathbb{R}^N, \mathcal{H}) into the bounded linear operators satisfying As(θ)Js=Dp(θ)A_s(\theta) J_s = Dp(\theta), such that |⟨As(θ)ei,p(η)⟩|≤C(m/N)log⁡m|\langle A_s(\theta) e_i, p(\eta)\rangle| \le C (m/N) \log m whenever ∥θ∥,∥η∥≤3m\|\theta\|, \|\eta\| \le 3\sqrt{m} and 1≤i≤N1 \le i \le N, while uniformly in θ∈ℝm\theta \in \mathbb{R}^m and 1≤s≤S1 \le s \le S the norms ∥p(θ)∥\|p(\theta)\|, ∥Dp(θ)∥op\|Dp(\theta)\|_{\mathrm{op}}, ∥D2p(θ)∥op\|D^2 p(\theta)\|_{\mathrm{op}}, ∥As(θ)∥op\|A_s(\theta)\|_{\mathrm{op}}, and ∥DAs(θ)∥op\|DA_s(\theta)\|_{\mathrm{op}} are at most m300m^{300}.

What Is Checked

The chart-extension claims are stated over the chart extension αg,B,q,k,ε\alpha_{g,B,q,k,\varepsilon}, the frame chart operator JuJ_u, and the residual operator I−JBI - JB for SS families of NN vectors gg in a real inner-product space EE, continuous linear maps BB from F=ℝNF = \mathbb{R}^N to EE, real scalars q,k,ε∈ℝq, k, \varepsilon \in \mathbb{R}, chart index s∈[S]s \in [S], and pair j∈[S]×[N]j \in [S] \times [N], with x,h=(hi)i∈[N]∈Fx, h = (h_i)_{i \in [N]} \in F and t∈ℝt \in \mathbb{R} in (7), under k≥0k \ge 0, ε>0\varepsilon > 0, and ∥(I−Jq−1gsBs)x∥≤R\|(I - J_{q^{-1} g_s} B_s)x\| \le R for R∈ℝR \in \mathbb{R} in (8), and under q>0q > 0, 0≤k≤10 \le k \le 1, ε>0\varepsilon > 0, ∑i∈[N]|hi|≤H\sum_{i \in [N]} |h_i| \le H for H∈ℝH \in \mathbb{R}, ∥gs,i∥≤2q\|g_{s,i}\| \le 2q, and ∥Bsei∥≤4k\|B_s e_i\| \le 4k in (9); the kernel-segment claim (10) is stated over the Hilbert kernel Kg,q,ε,a,bK_{g,q,\varepsilon,a,b}, the kernel pairing Φg,q,ε\Phi_{g,q,\varepsilon}, and the kernel pairing differential Ψg,q,ε\Psi_{g,q,\varepsilon} for a finite index set ι\iota, a vector family g=(gi)i∈ιg = (g_i)_{i \in \iota} in EE, real scalars q,ε,a,b,C,D∈ℝq, \varepsilon, a, b, C, D \in \mathbb{R}, vectors θ,η,v∈E\theta, \eta, v \in E, and coefficient functions α,β,δ:ι→ℝ\alpha, \beta, \delta : \iota \to \mathbb{R} under 0<a≤b0 < a \le b, D≥0D \ge 0, and |Ψg,q,ε(σ;θ+tv,η,v,α+tδ,β,δ)|≤C/σ+D/σ2|\Psi_{g,q,\varepsilon}(\sigma; \theta + t v, \eta, v, \alpha + t\delta, \beta, \delta)| \le C/\sigma + D/\sigma^2 for all t∈[0,1]t \in [0, 1] and σ∈[a,b]\sigma \in [a, b].