Gaussian embedding stability for finite persistence measures

Weighted Gaussian embeddings vary continuously under native first-order boundary transport. The inputs are positive finite Borel measures on the strict birth–death half-plane, including unequal masses and measures with nonatomic or noncompact support. Finite diagrams are recovered by their atomic measures.

Measures, weights and distances

Write H = {(b,d) ∈ ℝ² : b < d} and δ(b,d) = (d−b)/2. The ground distance is the maximum-coordinate distance on birth–death pairs. A boundary plan has the specified interior marginals and free diagonal marginals; its first-order cost is the integral of this ground distance. The native distance BT₁ is the infimum over all such plans, with values in [0,∞]. See the boundary-plan construction.

Let w be nonnegative, bounded by B and A-Lipschitz in maximum-norm birth–persistence coordinates, with w(b,0) = 0 and A,B ≥ 0. Put T(b,d) = (b,d−b). For nonzero bandwidth σ, the full-plane Gaussian kernel and weighted feature are

\[ k_\sigma(p,q)=\exp\!\left(-\frac{\|p-q\|_2^2}{2\sigma^2}\right), \qquad F_\sigma(p)=w(Tp)k_\sigma(\cdot,p). \]

The norm below is the norm of this kernel's reproducing kernel Hilbert space. Its sections have norm one, so ‖Fσ(p)‖ ≤ B. For every finite original measure μ, the weighted feature is Bochner integrable and

\[ \Phi_\sigma(\mu)=\int_H F_\sigma(p)\,d\mu(p). \]

This is the existing weightedPersistenceEmbedding σ (upperWeight w) μ. Interior and UpperPoint are the same strict-half-plane subtype; the bridge records their measurable equivalence and the integral identity explicitly.

Extended and real estimates

Set C = 2A + √2 B/|σ|. For arbitrary finite original μ and ν,

\[ \operatorname{ofReal}\|\Phi_\sigma(\mu)-\Phi_\sigma(\nu)\| \le \operatorname{ofReal}(C)\,\mathrm{BT}_1(\mu,\nu). \]

The inequality is in ENNReal, so the right side may be infinite. Neither integrability of the embedding nor this estimate needs a diagonal or ambient first moment. If ∫δ dμ and ∫δ dν are finite, the existing transport finiteness theorem proves BT₁ finite, and conversion to real numbers gives

\[ \|\Phi_\sigma(\mu)-\Phi_\sigma(\nu)\| \le C\,\operatorname{toReal}(\mathrm{BT}_1(\mu,\nu)). \]

The Lean statements also allow σ = 0 under the library's totalized division convention: the kernel is then constant one, and the same proof retains the weight term 2A. The usual nonzero-bandwidth theorem is a specialization.

Direct boundary-plan proof

The existing pointwise estimates give ‖Fσ(p)−Fσ(q)‖ ≤ C‖p−q‖∞ and ‖Fσ(p)‖ ≤ Cδ(p). The feature vanishes on the diagonal, so it extends by zero through the collapsed boundary. For any admissible plan π, its interior marginal equations identify the two endpoint integrals with Φσ(μ) and Φσ(ν). Free diagonal mass contributes zero. Every admissible plan is finite, which makes both bounded endpoint features integrable even when its transport cost is infinite.

The norm-of-integral inequality and the pointwise bound give ‖Φσ(μ)−Φσ(ν)‖ ≤ C∫‖p−q‖∞ dπ in extended nonnegative values. Taking the infimum over the actual admissible plans gives the native estimate. The zero-constant case uses an explicit admissible deletion/insertion plan before passing to the infimum. The argument requires no optimizer or probability normalization.

Atomic inputs and approximation

For a finite indexed family with nonnegative real masses mᵢ at points pᵢ,

\[ \Phi_\sigma\!\left(\sum_i m_i\delta_{p_i}\right) =\sum_i m_iF_\sigma(p_i). \]

Zero masses disappear and coincident coordinates add their masses. The native finite weighted diagram uses this same atomic measure, and its finite transport optimum agrees with the all-plan value. Thus the stability bound applies directly to fractional weighted diagrams and to the existing unit-occurrence embedding.

For finite original μ with finite first diagonal moment, let Dₙ be the existing quantizedDiagram μ n. Each atomic measure μDₙ has exactly μ's original mass, and BT₁(μ,μDₙ) tends to zero. The new estimate therefore proves Φσ(μDₙ) → Φσ(μ) in the actual RKHS norm. These are the native same-mass quantizations, with no new choice of approximants.

The consumers include a nonatomic band of bars (t,t+2), t ∈ [0,1], and the countable family (2ⁿ⁺²,2ⁿ⁺²+2) with masses 2⁻ⁿ⁻¹. The latter has mass one and diagonal moment one, while its ambient first moment diverges. It demonstrates why the real bound and approximation theorem use diagonal moments.

Sources and scope

Kusano–Fukumizu–Hiraoka, equation (3), Proposition 10 and Lemma 16, supplies the weighted Gaussian embedding and finite-diagram stability context. Here Lk = √2/|σ|, Bk = 1, B₂ = B and L₂ = 2A. The direct theorem for arbitrary finite original measures and native boundary transport is a separately proved library extension.

Boundary transport originates with Figalli–Gigli in bounded Euclidean domains with quadratic cost. Divol–Lacombe develops the persistence-measure setting; the upper-half-plane, general-exponent and maximum-norm formulation is distinct from Figalli–Gigli's original setting. The independent ordinary-Wasserstein comparison and corrected-article source comparison retain their separate boundary.

Adams et al. studies persistence images and pixel norms. Obayashi–Hiraoka studies linear prediction and inverse analysis. Neither supplies the RKHS stability claim proved here. Fixed-pixel stability, infinite-original-mass transport, essential bars and exponents above one lie outside this result.