ProductGeometricKernel Overview

• ProductGeometricKernel is a composite construct that forms the pointwise product of geometric kernels on spaces like manifolds, meshes, and graphs to enable structured kernel models.
• Its spectral characterization using orthogonal polynomial and Fourier expansions guarantees strict positive definiteness and provides practical recipes for kernel construction.
• Implemented in packages like GeometricKernels, it supports applications in spatial statistics, Gaussian processes, and spatio-temporal modeling through efficient feature map computations.

The ProductGeometricKernel is a mathematical and computational construct representing the pointwise product of two or more geometric kernels, enabling composite kernel models on product spaces such as manifolds, meshes, and graphs. Product kernels generalize to key classes of compact homogeneous spaces, where their structure and positive definiteness properties can be fully characterized through spectral criteria involving orthogonal polynomial expansions. In practical terms, the ProductGeometricKernel formalizes the interplay of individual geometric, spectral, or graph-based similarities, with substantive applications in areas including spatial statistics, uncertainty quantification, and geometric machine learning.

1. Mathematical Definition and Spectral Characterization

Let $k_1$ and $k_2$ be positive semi-definite kernels on spaces $X_1$ and $X_2$, respectively: $$k_1 : X_1 \times X_1 \to \mathbb{R}, \quad k_2 : X_2 \times X_2 \to \mathbb{R}.$$ Their product kernel on $X_1 \times X_2$ is defined as: $$k_{\mathrm{prod}}\bigl((x_1,x_2), (y_1, y_2)\bigr) = k_1(x_1, y_1) \cdot k_2(x_2, y_2).$$ If both kernels act on the same space, this reduces to $k_{\mathrm{prod}}(x, y) = k_1(x, y) k_2(x, y)$ (Mostowsky et al., 2024).

When each factor kernel admits a Mercer expansion, the product kernel inherits a double expansion: $$k_{\mathrm{prod}}(x, y) = \sum_{n=0}^\infty \sum_{m=0}^\infty \bigl(\lambda^{(1)}_n \lambda^{(2)}_m\bigr) \bigl[\varphi^{(1)}_n(x_1) \varphi^{(2)}_m(x_2)\bigr] \bigl[\varphi^{(1)}_n(y_1)\varphi^{(2)}_m(y_2)\bigr].$$ The associated feature map is the tensor product of the factors' feature maps: $$\phi_{\mathrm{prod}}(x_1, x_2) = \phi_1(x_1) \otimes \phi_2(x_2)$$.

For specific homogeneous spaces, this structure further refines. On the torus $S^1 \times S^1$, every continuous, isotropic, positive definite kernel admits a double Fourier expansion: $$K\bigl((\theta_1, \phi_1), (\theta_2, \phi_2)\bigr) = \sum_{m, n \in \mathbb{Z}} a_{m, n} e^{i[m(\theta_1 - \theta_2) + n(\phi_1 - \phi_2)]},$$ with required symmetry and nonnegativity for the coefficients: $$a_{m, n} = a_{-m, n} = a_{m, -n} = a_{-m, -n} \geq 0$$, and $\sum_{m, n} a_{m, n} < \infty$ (Guella et al., 2015).

2. Strictly Positive Definiteness of Product Kernels

The strict positive definiteness (SPD) of a product geometric kernel reduces to combinatorial properties of the support of its spectral expansion coefficients. For compact two-point homogeneous spaces, let the kernels' isotropic parts admit orthogonal polynomial expansions in Jacobi polynomials: $$f(t) = \sum_{k \ge 0} a_k(f) P_k^{(\alpha, \beta)}(t), \quad g(t) = \sum_{l \ge 0} a_l(g) P_l^{(\alpha, \beta)}(t).$$ Their product has coefficients supported at index sums or at specific linear combinations, depending on the manifold.

The necessary and sufficient conditions for SPD are as follows (Bonfim et al., 2018):

• On $S^1$ (Chebyshev): The set $\{\pm k \pm l : a_k(f) a_l(g) > 0\}$ must meet every full arithmetic progression in $\mathbb{Z}$.
• On $S^d,\,d\ge2$ (Gegenbauer): The set $\{k+l : a_k(f) a_l(g) > 0\}$ must have infinitely many even and infinitely many odd elements.
• On other projective spaces or the Cayley plane (Jacobi): The sum-index set must be infinite.

For the torus $S^1 \times S^1$, the extension of Schoenberg’s theorem asserts SPD if the support $I = \{(m, n) : a_{m, n} > 0\}$ “intersects every translate of every lattice in $\mathbb{Z}^2$” (Guella et al., 2015). This yields practical recipes for kernel construction.

3. Computational Implementation in GeometricKernels

Within the GeometricKernels package, the ProductGeometricKernel class enables the assembly of product kernels from multiple factor GeometricKernel objects. The kernel’s matrix evaluation computes the pointwise (Hadamard) product of the factor matrices:

K1 = kernel1.K(params1, x, x′) K2 = kernel2.K(params2, x, x′) return K1 * K2 # elementwise multiply

Spectral feature representations combine the factors’ feature maps via row-wise Kronecker products, compatible with Laplacian eigenfunction, Gegenbauer, or random Fourier features.

The implementation supports backend-agnostic automatic differentiation (NumPy, JAX, PyTorch, TensorFlow) and utilizes multiple dispatch to select the correct array operations. The computational complexity is $O(N^2)$ for matrix-matrix or matrix-vector evaluations (per factor), plus $O(N^2)$ for the final elementwise multiplication. Feature map sampling for $N$ points and map widths $\ell_1, \ell_2$ scales as $O(N\ell_1 + N\ell_2)$ for the factors, and $O(N\ell_1\ell_2)$ for the combined Kronecker feature map (Mostowsky et al., 2024).

4. Canonical Examples

Several canonical constructions illustrate the product kernel on specific spaces:

• Product of von Mises kernels on the torus: Each S¹ factor uses a von Mises expansion, leading to strictly positive definite kernels with parameter-controlled smoothness. The double index support I = $\mathbb{Z} \times \mathbb{Z}$ ensures SPD (Guella et al., 2015).
• Heat kernels on S¹ × S¹: Exponential decay in the double Fourier coefficients creates SPD kernels whose smoothness increases with the decay rates σ, τ.
• Product of Matérn or heat kernels on spheres, graphs, and meshes: Useful for modeling covariance with separable spatial, temporal, or structural dependencies (Mostowsky et al., 2024).
• Space–Time Products: For instance, combining a group kernel over $G$ (e.g., time) with a spherical kernel yields composite models with SPD criteria determined by the cross-coefficient support sets (Bonfim et al., 2018).

5. Applications: Modelling, Learning, and Approximation

Product geometric kernels are foundational in applications requiring rich, separable structure:

• Gaussian Processes and Kernel Methods: SPD guarantees nonsingular covariance and enables fast, spectral approximations in Gaussian process regression, classification, and kernel principal component analysis. On periodic or toroidal domains, product kernels capture the correct geometry for angular or temporal data (Guella et al., 2015).
• Spatio-Temporal Modeling: Product kernels combine spatial and temporal components (e.g., Matérn on a mesh times Matérn on a circle/time), resulting in separable models fit for kriging and spatial statistics (Mostowsky et al., 2024).
• Multi-Modal Data Fusion: By multiplying a kernel encoding spatial geometry with one encoding auxiliary features or relational structure, ProductGeometricKernel enables cross-domain covariance modeling (e.g., mesh & graph, positions & attributes).
• Approximation Theory: In interpolation and radial basis function schemes on manifolds, SPD kernels yield uniquely solvable collocation systems with native space error estimates (Bonfim et al., 2018). On S¹ × S¹, product kernel machinery supports function reconstruction for toroidal/surface data (Guella et al., 2015).

6. Interface and Customization in Practice

The ProductGeometricKernel class exposes APIs for initialization, parameter management, and feature map computation. Initialization aggregates factor kernel hyperparameters:

params = prod_kernel.init_params() params["0"]["lengthscale"] = ... params["1"]["lengthscale"] = ...

Normalization (scaling diagonal entries to one) and factor-by-factor hyperparameter adjustment are supported. Example deployments span spheres, graphs, and their cross-products, allowing compositional models reflecting application geometry (Mostowsky et al., 2024).

7. Theoretical and Practical Significance

ProductGeometricKernel offers a mathematically controlled route to constructing strictly positive definite, continuous, isotropic kernels on composite geometric domains. Its spectral characterization—via summing or combining spectral supports—translates to easily verifiable, implementable conditions for invertibility and well-posedness in statistical and approximation schemes (Bonfim et al., 2018). The mechanism also provides a palette for blending smoothness, locality, and attribute sensitivity, and supports automatic relevance determination via independent control of factor hyperparameters (Mostowsky et al., 2024).

Such composite kernels thus serve as indispensable primitives in modern geometric machine learning, signal processing on manifolds, and high-dimensional approximation tasks on structured domains.