Source-linked AI summary
Scattered Data Interpolation on Embedded Submanifolds with Restricted Positive Definite Kernels: Sobolev Error Estimates
Edward Fuselier, Grady Wright
TL;DR
Intrinsic manifold kernels may be impractical when the manifold is incompletely known or difficult to describe. The paper instead restricts positive-definite ambient kernels, characterizes their native spaces, and derives interpolation error estimates; numerical experiments report agreement with the predicted estimates.
Problem
Intrinsic kernels may be impractical to construct when the manifold is incompletely known or lacks a manageable mathematical description.
Source: §1 · p. 2Method
The paper restricts positive-definite ambient kernels to embedded manifolds, characterizes finite-smoothness native spaces, and derives Sobolev-type interpolation error estimates.
Source: §1 · p. 2Results
The extension map is an isometry between the restricted and ambient native spaces, and numerical errors show good agreement with predicted estimates in the reported experiments.
Source: §3 · p. 6, §5 · p. 16Takeaways & Limitations
Simple ambient-space kernels can be theoretically and numerically viable for approximating functions defined on manifolds.
Source: §1 · p. 2The theory assumes smooth, connected, compact manifolds without boundary; node-measure discrepancy may be large unless the node set is sufficiently dense.
Source: §2 · p. 3, §4.1 · p. 8
Abstract
from arXiv · showhide
In this paper we investigate the approximation properties of kernel interpolants on manifolds. The kernels we consider will be obtained by the restriction of positive definite kernels on $\R^d$, such as radial basis functions (RBFs), to a smooth, compact embedded submanifold $\M\subset \R^d$. For restricted kernels having finite smoothness, we provide a complete characterization of the native space on $\M$. After this and some preliminary setup, we present Sobolev-type error estimates for the interpolation problem. Numerical results verifying the theory are also presented for a one-dimensional curve embedded in $\R^3$ and a two-dimensional torus.
1 Introduction
The paper studies interpolation on embedded manifolds using restrictions of ambient positive-definite kernels, motivated by cases where an intrinsic kernel is impractical to construct. It outlines a characterization of restricted-kernel native spaces, Sobolev error estimates, and numerical verification.
1 Introduction
- For scattered nodes, interpolation seeks a shifted-kernel fit matching target values, with coefficients obtained by inverting the positive-definite kernel’s Gram matrix.The paper focuses on traditional value interpolation rather than more general integral or derivative data.Source: §1 · p. 1
- Restricting a positive-definite ambient kernel to an embedded manifold preserves positive definiteness and yields a kernel suitable for manifold interpolation.The paper investigates whether simple ambient-space kernels can approximate functions defined on manifolds.Source: §1 · p. 2
- Intrinsic manifold kernels may be impractical when the manifold is incompletely known or lacks a manageable mathematical description, as in geometry, learning, and surface-PDE applications.Examples include geometric objects represented by meshes or point clouds, and high-dimensional data assumed to lie on a virtually unknown low-dimensional manifold.Source: §1 · p. 2
- The restricted-kernel method is mesh-free, accommodates meshes, level sets, or point clouds, supports smooth interpolant derivatives, and uses extrinsic coordinates to bypass coordinate singularities.The cited example of coordinate singularities is the unit sphere S2.Source: §1 · p. 2
- Earlier studies of restricted kernels focused on spheres; this paper addresses generally nonzonal restrictions using Madych–Nelson variational theory and Narcowich et al.’s error analysis.For radial ambient kernels, restriction to the sphere is zonal and permits Laplace-Beltrami eigenfunction expansions.Source: §1 · p. 2
- The paper plans to characterize native spaces for restricted kernels, derive interpolation error estimates for smooth and rough targets, and verify the estimates numerically.The numerical examples are described as verification of the paper’s two main error estimates.Source: §1 · p. 2
2 Notation and Preliminaries
This section fixes the smooth, compact, boundaryless embedded-manifold setting and establishes the Sobolev, trace, and kernel-native-space tools used later. It also states how finite-smoothness kernels and their interpolants relate to familiar Sobolev spaces.
2 Notation and Preliminaries
- The paper restricts attention to smooth, connected, compact manifolds without boundary.Source: §2 · p. 3
2.1 Embedded Submanifolds in Rd
- An embedded submanifold of Rd has the subspace topology, and its inclusion into Rd is an immersion.Source: §2.1 · p. 3
2.1.1 Slice Charts
- Slice charts use ambient-space maps to straighten an embedded manifold locally into a k-dimensional slice.Source: §2.1.1 · p. 3
2.1.2 Distances on Embedded Submanifolds
- The manifold distance between two points is the infimum of lengths of piecewise smooth curves in the manifold joining them.Source: §2.1.2 · p. 3
2.1.3 Tubular Neighborhoods
- A compact smooth embedded manifold has a tubular neighborhood with a smooth retraction onto the manifold that is the identity there.The neighborhood radius can be reduced so the retraction's domain is compact.Source: §2.1.3 · p. 4
2.2 Sobolev Spaces
- Sobolev spaces on manifolds are defined through chart-based projections and partition-of-unity norms, whose equivalence does not depend on those choices.The section also introduces fractional-order Sobolev spaces.Source: §2.2 · p. 4
2.3 Traces
- For a compact k-dimensional embedded manifold, the trace maps Hτ(Rd) onto Hτ−(d−k)/2(M), with a bounded right inverse independent of τ.Source: §2.3 · p. 5
2.4 Positive Definite Kernels and Native Spaces
- For kernels with Fourier transforms of algebraic decay, the native space is Hτ(Rd) with equivalent norms, and kernel interpolation is an orthogonal projection.The interpolant and its residual each have native-space norm no greater than the target's.Source: §2.4 · p. 5
3 Native Spaces for Restricted Kernels
The section characterizes the native space of a positive definite kernel restricted to an embedded manifold. It identifies that space with a Sobolev space whose order reflects the ambient-to-manifold dimension difference.
3 Native Spaces for Restricted Kernels
- The proof uses a variational characterization of native-space membership that avoids Fourier transforms.The approach is attributed to Madych and Nelson.Source: §3 · p. 6
- The extension operator from the restricted-kernel native space into the ambient native space preserves norms and extends functions off the manifold.Source: §3 · p. 6
- Continuity of the trace operator supplies the reverse inclusion by bounding the manifold Sobolev norm with the restricted-kernel native-space norm.Source: §3 · p. 7
- For φ satisfying condition (1), the restricted kernel's native space is Hτ−(d−k)/2(M), with equivalent norms.Source: §3 · pp. 6–7
4 Interpolation via Restricted Kernels
This section relates node distribution on an embedded manifold to chart-based and ambient-space measures, then uses those relationships to state Sobolev interpolation error estimates. The estimates cover native-space targets, smoother targets, and rougher targets, with nonuniformity affecting the rough-target bound.
4 Interpolation via Restricted Kernels
- The estimates use node measures across ambient, intrinsic, and chart-mapped spaces, motivating comparisons between these measures before deriving error bounds.The section first relates separation radii, then establishes chart-distance and mesh-norm relationships.Source: §4 · p. 7, §4.1 · p. 8
4.1 Node Measures on Submanifolds of Rd
- Intrinsic and ambient separation radii are equivalent for finite node sets on a smooth compact embedded submanifold of Rd.Theorem 6 also states that intrinsic manifold distance and ambient Euclidean distance are equivalent on M.Source: §4.1 · p. 8
- When hX,M is sufficiently small, the global density formed from chart-wise mesh norms is equivalent to the intrinsic mesh norm.The text also states that hX,M is bounded above by this chart-wise density.Source: §4.1 · p. 9
4.2 Interpolation Error Estimates
- For targets in Hs(M), Theorem 10 gives an interpolation error estimate when the intrinsic mesh norm hX,M is sufficiently small.The supplied theorem statement does not include the bound itself.Source: §4.2 · p. 10
- For very smooth targets, the spline-theory doubling trick gives Sobolev error estimates with doubled order in hX,M; functions with T^-1f in L2(M) have faster convergence.The text identifies T^-1f in L2(M) as a condition for smoother functions, but the supplied passages do not include the full estimate formulas.Source: §4.2 · p. 11
- For rougher targets f in Hβ(M), Theorem 14 gives an error estimate under small hX,M and the conditions s > β > k/2 and ⌊β⌋ > μ + k/2.The proof extends f to Rd and uses a band-limited interpolant.Source: §4.2 · p. 12
- For rough targets, nonuniform nodes can make the error bound large, whereas more-or-less uniform node sets keep ρX,M bounded and yield typical approximation rates.The mesh ratio is defined as ρX,M := hX,M/qX,M.Source: §4.2 · pp. 12–13
5 Numerical results
Numerical experiments on a curve and a torus in R3 test the paper’s interpolation estimates for targets inside and outside the kernel’s native space. Across both manifolds, the measured errors agree well with the predicted estimates, including the smooth-target doubling estimate.
Experimental setup
- Tests use a curve and a torus in R3 with near-minimum Riesz-energy node sets; mesh ratios remain roughly constant, and torus nodes are not aligned to vertices or lines.The plotted node sets use N = 100 on the curve and N = 1000 on the torus.Source: §5 · pp. 13–14, Figure 1 · p. 14
- The Wendland RBF has native space H4(R3), inducing H3(M1) and H3.5(M2); experiments fix support radius δ = 8/3 rather than study its selection.The kernel is compactly supported and has four continuous derivatives.Source: §5 · p. 14
Target functions
- Targets are linear combinations of Matérn kernels, with coefficients chosen to interpolate prescribed values at scattered points; fβ is not the comparison interpolant.The experiments use m = 25 points for M1 and m = 100 for M2.Source: §5 · p. 15
Native-space experiment
- For β = 4 targets, the computed relative errors on both manifolds agree well with the estimates from Theorem 10.The paper identifies these targets as suitable for Theorem 10 because the kernel’s ambient native space is H4(R3).Source: §5 · p. 16
Rough-target experiment
- For β = 3.5 targets rougher than the kernel’s native space, the predicted ℓ2 errors decrease like h2.5, and computed errors agree well with the estimates.The prediction assumes the mesh ratio is roughly constant, as in these experiments.Source: §5 · p. 16
Final smooth-target experiment
- For the C∞ target, measured ℓ2 errors on both manifolds agree well with Corollary 12’s estimates, while the results show no apparent reduction in the ℓ∞ error estimates.The predicted ℓ2 error decrease is h6.Source: §5 · p. 16