Gaussian Process Manifold Interpolation (GPMI)

• Gaussian Process Manifold Interpolation (GPMI) is a nonparametric framework that models and interpolates functions on manifolds while respecting geometric and topological constraints.
• It constructs specialized covariance kernels—using methods like heat kernels, spectral approximations, and extrinsic mappings—to adapt GP inference to curved spaces.
• GPMI enables advanced applications in shape analysis, Bayesian regression, sensor localization, and dynamical system parameter estimation with reliable uncertainty quantification.

Gaussian Process Manifold Interpolation (GPMI) is a framework for nonparametric regression and interpolation of functions, fields, or geometrical structures defined on data lying intrinsically on a manifold or a non-Euclidean domain. GPMI methodologies adapt classical Gaussian process (GP) inference to account for the manifold’s intrinsic geometry, topology, and smoothness properties, enabling uncertainty quantification and principled interpolation that respects geometric constraints. GPMI underpins modern algorithms for Bayesian regression, shape analysis, parameter estimation in dynamical systems, and probabilistic mapping on curved surfaces, with specialized models for spheres, Grassmannians, directed tori, and biological manifolds.

1. Theoretical Foundations and Problem Statement

GPMI considers data of the form $\{(x_i, y_i)\}_{i=1}^n$ where $x_i$ are locations on a manifold $M$ and $y_i = f(x_i) + \epsilon_i$, with $\epsilon_i \sim N(0, \sigma^2)$. The primary objective is to recover or interpolate the underlying function $f:M \to \mathbb{R}$ (or a vector-valued generalization) at unobserved locations on $M$, while quantifying uncertainty and maintaining respect for the manifold’s geometric structure.

On general Riemannian manifolds $(M, g)$, the prior covariance kernel is constructed using either:

• The heat kernel $$k_t(x, y) = \sum_{i=0}^\infty e^{-\lambda_i t}\; \phi_i(x)\; \phi_i(y)$$, where $\{\phi_i\}$ and $x_i$0 are eigenfunctions and eigenvalues of the Laplace–Beltrami operator $x_i$1 on $x_i$2 (Niu et al., 2018, Ye et al., 2020, Gao et al., 2018, Coveney et al., 2020).
• Kernels that are adapted to the manifold’s topology, such as hypertoroidal kernels or manifold-appropriate covariance functions for spheres, Grassmannians, or directed tori (Cao et al., 2023, Zhang et al., 2021, Lin et al., 2017).

Two primary paradigms exist:

1. Intrinsic GPMI: Direct construction of the kernel using the manifold’s geometry, often via the heat kernel or Brownian motion transition densities (Niu et al., 2018, Ye et al., 2020).
2. Extrinsic GPMI: Embedding $x_i$3 into an ambient Euclidean space and pulling back a standard kernel (e.g., squared exponential) via the embedding (Lin et al., 2017).

Both approaches yield tractable, positive semi-definite covariance matrices supporting posterior inference via conditioning.

2. Kernel Construction and Geometric Adaptation

The key technical advance in GPMI is in constructing covariance kernels $x_i$4 that encode the geometry of $x_i$5:

• Heat Kernel Approaches: On a compact manifold, the heat kernel $x_i$6 is the transition density of Brownian motion and solves $x_i$7. This kernel interpolates the squared-exponential as a special case when $x_i$8 but notably respects curvature, holes, and boundaries (Niu et al., 2018, Gao et al., 2018, Ye et al., 2020).
• Monte Carlo Estimation: In high dimension or complex domains, $x_i$9 is estimated via Brownian motion simulation, using ball or strip count estimators. For example, simulate $M$0 Brownian paths from $M$1, count the fraction arriving within a window of $M$2 after time $M$3, normalize by the window’s volume, and estimate $M$4. Strip estimators are exponentially more efficient on homogeneous manifolds (Ye et al., 2020).
• Spectral Approximations: For practical problems (e.g., heart atrium), a finite-rank eigenfunction expansion of the Laplacian is used, yielding a reduced-rank GP model with covariance $M$5, where $M$6 is a Matérn or analogous spectral density (Coveney et al., 2020).
• Specialized Manifold Kernels: For product manifolds (e.g., tori), kernels such as the hypertoroidal von Mises kernel generalize periodic and circular similarities, capturing cross-component correlations and guaranteeing positive definiteness on $M$7 (Cao et al., 2023).
• Extrinsic Kernels: For embedded manifolds (e.g., spheres, SPD matrices), a mapping $M$8 transforms the problem, and the kernel is composed as $M$9 (with $y_i = f(x_i) + \epsilon_i$0 typically quadratic-exponential) (Lin et al., 2017).

3. Gaussian Process Priors, Posterior Inference, and Interpolation

After assembling a covariance kernel, GPMI proceeds with Gaussian process prior and inference:

• Prior: $y_i = f(x_i) + \epsilon_i$1 on $y_i = f(x_i) + \epsilon_i$2.
• Observation Model: $y_i = f(x_i) + \epsilon_i$3, $y_i = f(x_i) + \epsilon_i$4.
• Posterior: For new $y_i = f(x_i) + \epsilon_i$5,

$y_i = f(x_i) + \epsilon_i$6

where $y_i = f(x_i) + \epsilon_i$7 and $y_i = f(x_i) + \epsilon_i$8 (Niu et al., 2018, Lin et al., 2017).

• Uncertainty Quantification: The posterior variance quantifies interpolation uncertainty, shrinking near data and reflecting information geometry (Gao et al., 2018, Coveney et al., 2020).
• Hyperparameter Selection: Marginal likelihood optimization or, where required (e.g., Grassmannian regression), manifold-aware cross-validation (e.g., LOOCV using the geodesic subspace distance) (Zhang et al., 2021).
• Computation: Reduced-rank and specialized Kronecker or block-matrix structures (e.g., ICM for multi-output on tori) are used for scalability (Cao et al., 2023, Coveney et al., 2020).

4. Algorithmic Variants and Specialized Frameworks

GPMI encompasses a variety of implementations and extensions suited for different settings:

Variant	Key Feature	Notable Application Domains
Heat-kernel intrinsic GP	Brownian-motion/heat kernel native to $y_i = f(x_i) + \epsilon_i$9	Arbitrary manifolds, shape analysis
Spectral low-rank GP	Laplacian eigenbasis expansion	Surface fields, biological mapping
Hypertoroidal/circular GP	Positive-definite via von Mises on tori	Directional data, localization, sensors
Extrinsic GP	Embedding then standard kernel on $\epsilon_i \sim N(0, \sigma^2)$0	SPD matrices, planar shapes, Grassmannians
Grassmannian GP	Joint GP in ambient, MACG-induced on $\epsilon_i \sim N(0, \sigma^2)$1	ROM, subspace regression
Tangent-chart GPMI	Local PCA + local GP in tangent chart	Manifold denoising, probabilistic recon.

• Active Learning on Manifolds: Adaptive selection of inputs via posterior variance or ALC criteria works efficiently in low-dimensional latent/embedded space, using neural autoencoders and GP regression in latent space (Cheng et al., 26 Jun 2025, Gao et al., 2018).
• Manifold-Constrained GPs in Dynamic Systems: For ODEs and DDEs, GPMI constrains GP draws so that state trajectories and their derivatives obey the governing equations at finite points, enabling parameter inference without numerical integration (Zhao et al., 2024).
• Multi-output and Coregionalization: Intrinsic coregionalization enables vector-valued outputs, e.g., sensor networks with multiple correlated quantities (Cao et al., 2023).
• Geometric Landmarking and Experimental Design: Sequential selection of points by maximizing GP posterior variance yields near-optimal geometric design (Kolmogorov $\epsilon_i \sim N(0, \sigma^2)$2-width-optimal rate) for manifold sampling (Gao et al., 2018).

5. Practical Implementation and Computational Considerations

• Monte Carlo and Spectral Estimation: Monte Carlo methods for approximating kernels scale as $\epsilon_i \sim N(0, \sigma^2)$3 for $\epsilon_i \sim N(0, \sigma^2)$4 samples and $\epsilon_i \sim N(0, \sigma^2)$5 BM paths but can be efficiently parallelized. For geometric Laplacians, sparse eigensolvers and mesh preprocessing are used (Ye et al., 2020, Coveney et al., 2020).
• Reduced-Rank and Efficiency: If using $\epsilon_i \sim N(0, \sigma^2)$6 basis functions, computational cost is $\epsilon_i \sim N(0, \sigma^2)$7, with prediction scaling in $\epsilon_i \sim N(0, \sigma^2)$8 per test (Coveney et al., 2020).
• Boundary and Topology Handling: Reflecting boundaries and holes are handled intrinsically through the kernel or mesh extension (Neumann boundaries), avoiding spurious smoothing across gaps or boundaries (Niu et al., 2018, Coveney et al., 2020).
• Hyperparameter Optimization: Marginal likelihood gradients are analytic for extrinsic and specialized kernels. For some manifold-induced models (Grassmannians), direct likelihood is singular; cross-validation on the manifold (e.g., Riemannian subspace distance) replaces it (Zhang et al., 2021).
• Uncertainty Propagation: For derivative-based quantities (e.g., conduction velocity), GP posterior samples are differentiated, and their statistics (mean, variance, percentiles) computed, supporting uncertainty-quantified downstream analyses (Coveney et al., 2020).

6. Applications and Empirical Results

• Cardiac Electrophysiology: GPMI on triangulated 2D atrial surfaces yields smooth, differentiable posterior fields for local activation time (LAT) and their gradients (conduction velocity, CV), propagating measurement and spatial uncertainty (nRMSE for LAT interpolation $\epsilon_i \sim N(0, \sigma^2)$9 with 250+ points, CV nRMSE $f:M \to \mathbb{R}$0–$f:M \to \mathbb{R}$1) (Coveney et al., 2020).
• Sensor Networks: HvM kernel-based GPMIs in AoA-based localization outperform both naïve and periodic-product kernels, achieving finer root-mean-squared absolute-position errors ($f:M \to \mathbb{R}$2m median RMSE vs $f:M \to \mathbb{R}$3–$f:M \to \mathbb{R}$4m) (Cao et al., 2023).
• Model Order Reduction: Grassmannian GPMI with analytic MACG posterior and LOOCV tuning yields subspace prediction errors on par or better than local bases and tangential interpolation, with scalable online evaluation (Zhang et al., 2021).
• Active Design: Uncertainty-driven landmark selection yields fill-distance rates $f:M \to \mathbb{R}$5, matching n-width oracle bounds and outperforming random placements in manifold coverage and reconstruction accuracy (Gao et al., 2018).
• Delay Differential Equation Inference: GPMI-style trajectory inference recovers DDE parameters and latent state trajectories efficiently, often surpassing collocation and solver-based filters (RMSE $f:M \to \mathbb{R}$6 for parameters in Hutchinson’s equation under sparse/noisy observation) (Zhao et al., 2024).

7. Limitations, Extensions, and Research Directions

• Smoothness and Extrapolation: GP smoothness assumptions can induce mean-reversion and suppress genuine discontinuities; extending GPMI to non-stationary or changepoint kernels is active research (Coveney et al., 2020).
• Computational Scalability: While reduced-rank and inducing-point techniques mitigate cubic scaling, large or high-dimensional manifolds (e.g., volumetric meshes, large sensor arrays) remain challenging (Niu et al., 2018, Ye et al., 2020).
• Boundary/Topology Sensitivity: Kernel construction on highly curved or multiply-connected manifolds requires care; naïve Euclidean-distance kernels may violate positive definiteness when restricted to $f:M \to \mathbb{R}$7 (Lin et al., 2017).
• Integration with Dynamical/Semi-supervised Models: Recent work embeds GPMI within Bayesian hierarchical models, latent-force systems, and joint semi-supervised learning pipelines, and explores uncertainty-aware manifold encoding (Coveney et al., 2020, Cheng et al., 26 Jun 2025).
• Future Directions: Anticipated advances include closed-form or more efficient kernel approximations on symmetric spaces, automated diffusion-time selection, active design integrated with GPMI uncertainty, and broader integration in spatiotemporal, physical, and biological modeling (Ye et al., 2020, Coveney et al., 2020, Cheng et al., 26 Jun 2025).