Spectral Convergence of the connection Laplacian from random samples

1. Introduction

• The paper proves spectral convergence of graph connection Laplacians to connection Laplacians on associated vector bundles, extending earlier pointwise convergence of VDM.It also proves convergence of a normalized graph Laplacian to the Laplace-Beltrami operator beyond the uniform-sampling, boundary-free setting of earlier work.Source: §1 · pp. 1–2
• Euclidean distances are sensitive to nuisance parameters, so the paper uses group-invariant distances between transformation orbits instead.Examples of nuisance variation include acquisition conditions in fMRI and rigid transformations in shape analysis.Source: §1 · pp. 1–2
• The principal-bundle framework treats the total space as the bundle, the orbit space as the base manifold, and each orbit as a fiber.It supplies a common mathematical language for LE, DM, ODM, and VDM.Source: §1 · p. 4
• The convergence results extend to independent non-uniform samples and manifolds with boundary, subject to mild technical conditions for non-uniform sampling.The paper also proves spectral convergence when tangent bundles must be estimated from the point cloud.Source: §1 · pp. 2, 4
• The framework connects convergence guarantees to vector diffusion distances, manifold orientability, and classification of cryo-EM projection directions.For manifold learning, VDD approximates geodesic distance between nearby points; the paper also establishes a symmetric Nash embedding result for orientable double covers [22, 23].Source: §1 · pp. 3–4

2. The Graph Connection Laplacian and Vector Diffusion Maps

• The graph connection Laplacian encodes both data affinities and edge transformations, with a particle carrying a vector status that changes as it moves through the graph.For q = 1 and identity group elements, it reduces to the ordinary graph Laplacian; otherwise, vector coordinates generally remain coupled.Source: §2 · p. 5
• VDM defines affinities from consistency among vector statuses propagated along paths, so the affinity reflects both path counts and agreement among their transformations.The construction averages rotational information over paths of length 2t and uses a squared Hilbert-Schmidt norm.Source: §2 · pp. 5–6
• The vector diffusion distance compares the VDM embeddings of graph nodes, providing a distance based on the connection information encoded in the graph.The paper also describes a degree-normalized VDM variant that maps the data into a Hilbert space.Source: §2 · p. 6
• Because the normalized connection matrix need not have nonnegative eigenvalues, the paper defines VDM using a squared Hilbert-Schmidt norm.The unnormalized graph connection Laplacian is positive semidefinite.Source: §2 · p. 6

3.1. Notations and Background of Differential Geometry.

• The setup takes a compact smooth manifold embedded in Euclidean space and equips it with the induced metric and volume measure.A principal bundle has a total space, base manifold, and group fiber; its associated vector bundle has fibers represented by R^q.Source: §3.1 · p. 7
• The bundle connection defines covariant differentiation by comparing sections through horizontal lifts and parallel displacement between fibers.In the frame-bundle example, this comparison changes coordinates between tangent spaces using their bases.Source: §3.1 · p. 8

3.2. Notations and Background of Numerical Finite Samples.

• For finite samples, the paper represents the sampled bundle as a direct sum of fibers and normalizes kernels using estimated sampling density to reduce non-uniform-density effects.It assumes independent samples from a density p on the manifold and specifies kernel functions and empirical quantities for the analysis.Source: §3.2 · pp. 8–10

4. Unifying VDM, ODM, LE and DM from the Principal Bundle viewpoint

• The principal-bundle structure unifies VDM, ODM, LE, and DM and provides a framework for studying their corresponding operators.Different associated bundles can provide different geometric or topological information; the tangent-bundle connection Laplacian used by VDM also reflects nearby geodesic distances and the 1-Betti number mixed with Ricci curvature.Source: §4.1 · p. 11, §4 · p. 11, §4.2 · p. 14

4.1. Principal Bundle Setup.

• The setup uses a smooth embedded manifold, a principal bundle with connection, and an orthogonal representation; sample points are independent draws from a density bounded above and below.Source: §4.1 · pp. 11–12

4.2. Unifying VDM, ODM, LE and DM under the manifold setup.

• ODM uses the orientation bundle with group Z2, whereas LE and DM use the trivial bundle and ignore edge connection groups.Source: §4.1 · p. 11, §4.2 · p. 14
• The graph connection Laplacian is built from edge weights and group-valued connection data, which relate to parallel transport and accumulate local covariant-derivative information.Source: §4.2 · pp. 12–13
• ODM analyzes higher eigenvectors for orientable manifolds and modified diffusion maps recover the orientable double cover for non-orientable manifolds [28, Section 2.3] [28, Section 3].Source: §4.2 · p. 13
• Appendix D proves that a smooth, closed non-orientable manifold has an orientable double cover embedded symmetrically in some Rp, supporting its visualization with modified diffusion maps [22, 23].Source: §4.2 · pp. 13–14

5.1. Pointwise Convergence.

• The integral operator approximates the identity and yields pointwise connection-Laplacian convergence, with separate estimates for points outside and inside a boundary neighborhood.The stated estimates depend on section regularity; interior and boundary error terms are treated separately.Source: §5.1 · p. 15
• Pointwise convergence of the sampled GCL requires stochastic fluctuation smaller than its O(h) signal; the corollaries give high-probability interior and boundary results under corresponding fluctuation conditions.For points in the boundary neighborhood, the stated fluctuation condition is o(h^1/2); the results hold with probability higher than 1 − O(1/n^2).Source: §5.1 · pp. 15–16
• Near the boundary, the discrete eigenvectors approximate eigenvector fields of the connection Laplacian with homogeneous Neumann boundary conditions.The boundary condition is ∇∂dX(x) = 0 for x ∈ ∂M.Source: §5.1 · p. 17

5.2. Spectral Convergence.

• For fixed i, heat-kernel and connection-Laplacian eigenvalues and associated eigenvectors converge along a bandwidth sequence, with eigenvectors converging in L2 in probability.The spectral theorems assume simple eigenvalues and 2/5 < γ < 1/2; when the boundary is empty, the γ condition can be ignored.Source: §5.2 · p. 18
• For fixed h, bT_h,α,n converges pointwise almost everywhere and compactly almost surely to T_h,α as n→∞, yielding almost-sure spectral convergence.Source: §5.2 · pp. 22, 24
• In the boundary-free case, as h→0, h^-1(T_{h,1}−1)X spectrally converges to ∇²X; for t>0, T_{h,1}^{t/h} converges to e^(t∇²).Source: §5.2 · p. 24
• With a nonempty boundary, the proof incorporates the boundary effect and Neumann condition; it reaches the same conclusion when γ>2/5 under a bandwidth condition.Source: §5.2 · p. 25
• Putting the steps together concludes the theorem’s convergence in probability; the proof for Theorem 5.5 is the same and is omitted.Source: §5.2 · p. 26

6. Extract more Topological/Geometric Information from a Point Cloud

• Local PCA estimates tangent-plane bases, which are used to estimate the frame bundle and parallel transport before constructing the graph connection Laplacian.The parallel-transport estimate approximates transport between nearby tangent planes.Source: §6 · pp. 26–27
• For α = 1, the reconstructed operator corresponds to the connection Laplacian on the tangent bundle, while accumulated estimation errors must be controlled for spectral convergence.The proof controls errors from estimating tangent planes and parallel transport; under its stated events, frame-bundle estimation contributes a negligible error on a suitable subspace.Source: §6 · pp. 27, 28, 30
• Along a sequence of bandwidths, the estimated heat-kernel eigenvalues and eigenvector fields converge in probability to those of the connection-Laplacian heat kernel.The theorem assumes simple eigenvalues and uses estimated parallel transport to construct the graph connection Laplacian.Source: §6 · pp. 27–28
• Along a sequence of bandwidths, the estimated connection-Laplacian eigenvalues and eigenvector fields converge in probability to their continuum counterparts.The theorem applies to the graph connection Laplacian built using estimated parallel transport.Source: §6 · p. 28

Appendix A. An Introduction to Principal Bundle

• A principal bundle comprises a total space with a group action, whose orbit space is the base manifold.The group is the structure group, and the projection maps points in the total space to the base.Source: Appendix A · p. 32
• The frame bundle collects tangent-space bases, while the orientation bundle records the possible orientations at each point.The appendix also uses the trivial principal bundle to encompass graph Laplacian and diffusion-map cases.Source: Appendix A · pp. 32–33
• An associated vector bundle is formed from a principal bundle and a group representation; the tangent bundle is one such example.Changes of frame induce changes in tangent-vector coordinates.Source: Appendix A · p. 33
• A connection chooses a G-invariant complement to the vertical bundle, enabling parallel displacement and defining a covariant derivative.The connection Laplacian is then defined from the covariant derivative and is self-adjoint under the stated setup.Source: Appendix A · pp. 34, 35, 37

Appendix B. [Proof of Theorem 5.2]

• Near the boundary, kernel asymmetry means the proof does not expect a second-order term in the approximation.The boundary analysis accounts for the geometry of the domain to understand the first-order term.Source: Appendix B · p. 38
• Away from the boundary, the kernel expansion has error terms whose bounds depend on the function and its derivatives through order four.The stated regularity is F ∈ C4(M\Cx) ∩ L∞(M).Source: Appendix B · p. 38
• For points near the boundary, the expansion gives an O(h^2γ) error depending on the function and its first two derivatives.The proof derives this boundary-case order after expanding the numerator and denominator of the kernel operator.Source: Appendix B · pp. 42–43

Appendix C. [Proof of Theorem 5.3]

• The stochastic analysis generalizes [29, Theorem B.3] to principal bundles and uses large-deviation bounds to control the second-order term of size h.Bernstein’s inequality is applied with deviations much smaller than h.Source: Appendix C · p. 43
• Under Assumptions 4.1, 4.2, and 4.3, the stated kernel density estimation result holds for all sample points i = 1, ..., n with probability higher than 1 − O(1/n^2).Source: Appendix C · p. 43
• Near the boundary, the first-order term does not cancel, changing the variance order and requiring α=o(h^1/2) for the large-deviation bound.The variance is O(h^−(d/2−1/2)) near the boundary, rather than O(h^−(d/2−1)) away from it.Source: Appendix C · pp. 46, 48
• With probability higher than 1−O(1/n^2), the proof controls the relevant fluctuations across all sample points, including the boundary case.The argument combines high-probability events and handles α=0 separately.Source: Appendix C · pp. 47–48

Appendix D. Symmetric Isometric Embedding

• Odd Laplace eigenfunctions yield an embedding of M into ℓ2, but this embedding is neither isometric nor finite-dimensional.The argument establishes continuity, injectivity, and immersion before concluding that the map is an embedding.Source: Appendix D · p. 49
• A symmetry-preserving construction first produces a finite-dimensional symmetric embedding, then augments it so its first and second derivatives are linearly independent.That derivative independence supports the subsequent local metric corrections.Source: Appendix D · p. 52
• After scaling the preliminary embedding so its induced metric is smaller than g, the proof uses local corrections and a contraction argument to construct an isometric embedding.The contraction argument yields a local solution, which the proof uses to complete the patch.Source: Appendix D · pp. 52, 55, 56
• Every smooth closed manifold with a free isometric Z2 action admits a smooth symmetric isometric embedding into some finite-dimensional Euclidean space.The construction uses a large-λ symmetric auxiliary map and a symmetric correction to obtain the isometric embedding.Source: Appendix D · p. 54

• The perturbation preserves symmetry because the local quantities are unchanged under v → −v, and finitely many patches can be handled by choosing ε sufficiently small.After perturbing each patch, the modified embedding remains symmetric.Source: Appendix D · p. 56
• The appendix proves existence but does not determine the optimal dimension of the embedded Euclidean space.How to account for symmetry when finding an optimal isometric embedding is left for future work.Source: Appendix D · p. 56
• Every smooth, closed non-orientable manifold has an orientable double covering embedded symmetrically in some Euclidean space.The proof applies Theorem D.5 to the double covering, which has an isometric free Z2 action.Source: Appendix D · p. 56