Kernel Discretization in A-MWGraD

• Kernel-based discretization is a mesh-free numerical method that uses positive-definite kernels and reproducing kernel Hilbert spaces to approximate PDEs in the A-MWGraD framework.
• It employs particle-based estimators and tensor-product kernels to discretize scalar, vector, and matrix-valued operators, ensuring convergence and computational tractability.
• The approach integrates acceleration techniques like Nesterov updates, while preserving structural properties such as Hermitian symmetry and divergence, crucial for robust multi-objective optimization.

Kernel-based discretization for A-MWGraD refers to the principled use of positive-definite kernels to approximate and discretize the dynamics or equations arising in the Accelerated Multiple Wasserstein Gradient Decomposition (A-MWGraD) framework. This methodology provides a flexible, mesh-free means for numerically solving PDEs and evolution equations where the solution is either scalar, vector, or matrix-valued, as in contraction metrics or multi-objective Wasserstein flows. This approach exploits the structure of reproducing kernel Hilbert spaces (RKHS), particle methods, and tensor-product kernels to ensure convergence, computational tractability, and compatibility with acceleration and multi-objective optimization requirements (Nguyen et al., 27 Jan 2026, &&&1&&&).

1. Overview of A-MWGraD and the Need for Kernel-based Discretization

A-MWGraD (Accelerated Multiple Wasserstein Gradient Descent) generalizes Wasserstein gradient flows to handle simultaneous multi-objective optimization in distribution space, introducing acceleration in the sense of Nesterov. The continuous-time flow is given by

$\left\{ \begin{aligned} &\dot\rho_t + \nabla\cdot(\rho_t\,\nabla\Phi_t) = 0, \[6pt] &\dot\Phi_t + \alpha_t\,\Phi_t + \tfrac12\|\nabla\Phi_t\|^2 + \mathrm{proj}_{\mathcal C(\rho_t),\rho_t}[0] = 0, \end{aligned} \right.$

with $\rho_t$ a measure on $\mathcal X$, $\Phi_t$ a potential, complex projection/prox operators over first variations, and $\alpha_t$ a time-varying damping (Nguyen et al., 27 Jan 2026). Discretization in practice hinges on approximating $\nabla\log\rho(\cdot)$, the coupling between measure and test functions, and, when working with PDEs or operator flows, the derivative and bilinear structure of underlying spaces. Kernel-based methods address the need for differentiable, positive-definite smoothing and mesh-free approximation in these settings (Giesl et al., 2017).

2. Kernel-based Discretization Principles

Kernel discretization replaces or augments classical grid-based or finite-element representations by using a set of centers $\{x_j\}$ and a scalar or tensor-valued kernel $K_h(x, y)$ (typically Gaussian or compactly supported) to interpolate functionals, gradients, or matrix-valued fields. For scalar or vector fields, the smoothing operator is: $$(\mathcal K_h\phi)(x) = \int K_h(x,y)\,\phi(y)\,dy.$$ For particles, the integral is approximated by empirical sums over samples; for matrix-valued equations, fourth-order tensor kernels are used (Giesl et al., 2017). Kernel smoothers enforce positive-definiteness, locality, and allow natural interpolation of gradients and first variations required in A-MWGraD or operator versions thereof.

3. Kernel Discretization in Particle-based A-MWGraD

For multi-objective Wasserstein optimization (e.g., multi-target sampling), $F_k(\rho) = KL(\rho\|\pi_k)$, with $$\nabla\delta_\rho F_k[\rho](x) = \nabla f_k(x) + \nabla\log\rho(x)$$. Kernel-based estimates of $\nabla\log\rho(x)$, essential for SVGD and blob methods, entail:

• SVGD estimator:

$$\bar\Delta^{(n)}_k(x_i) = \frac{1}{m}\sum_{j=1}^m \left[ K_h(x_i, x_j)\nabla f_k(x_j) - \nabla_{x_j} K_h(x_i, x_j) \right]$$

• Blob estimator uses normalization by local kernel sum.

The bandwidth $h$ is usually set by the median heuristic: $h^2 = \mathrm{median}\{\|x_i-x_j\|^2: i<j\}$ (Nguyen et al., 27 Jan 2026). Discrete A-MWGraD iterates use these kernel gradients, combine them with convex weights, and update particles with Nesterov-style extrapolation/correction.

4. Kernel-based Discretization for Matrix-valued PDEs

When A-MWGraD is posed in operator or PDE settings (e.g., contraction-metric, photonic crystal eigenproblems), kernel-based discretization is implemented via:

• Tensor-product kernels: $K(x, y)_{ijkl} = \phi(x, y)\delta_{ik}\delta_{jl}$, $\phi$ a scalar positive-definite kernel.
• Approximation ansatz:

$M_h(x) = \sum_{j=1}^N K(x, x_j) C_j,$

where $C_j$ are symmetric coefficient matrices (Giesl et al., 2017).

• Collocation: Enforcing the PDE $A(M_h)(x_i) = -C(x_i)$ at centers yields block-linear systems for $\{C_j\}$.

Error estimates of the form $$\|M - M_h\|_{L^\infty}\leq C h^{\sigma - 1 - n/2}\|M\|_{H^\sigma}$$ are established, where $h$ is the fill distance and $\sigma$ the kernel's smoothness (Giesl et al., 2017).

5. Structure-preserving and Acceleration-compatible Discretization

In Maxwell/eigenvector problems and applications like photonic crystals, kernel-based discretization is adapted to preserve Hermitian positive definiteness (HPD), block symmetry, and discrete analogues of divergence and curl via spectral and block circulant operators. The full discretization, when linked to A-MWGraD (as in the Algebraic Maxwell–Weighted Gradient Decomposition (A-MWGraD)), implements:

• Weighting and transfer via indicator diagonals and symmetrized operators $S_{ij}$;
• Penalty enforcement for null-space removal;
• GPU-acceleration via FFT-friendly circulant representations, enabling $O(N^3\log N)$ per-iteration scaling for high-dimensional systems (Jin et al., 21 Nov 2025).

When adapting to A-MWGraD, one replaces differential/curl operators and permittivity with their weighted versions, applies kernel-based transfer for cross-degrees of freedom, and retains the DFT-based matrix-free structure. The $\mathcal O(1)$ penalty parameter rule and HPD proofs generalize directly (Jin et al., 21 Nov 2025).

6. Practical Algorithms and Parameter Choices

Parameter selection in kernel-based A-MWGraD is informed by stability and accuracy analyses:

• Step size $\eta$: sweep over $10^{-3}$--$10^{-1}$; $\eta\sim10^{-3}$–$10^{-2}$ is robust in trials.
• Bandwidth $h$: always set by the median pairwise distance among particles.
• For PDEs: kernel support should include at least $2$–$4$ nearest neighbors.
• Computationally, kernel matrix sums cost $O(m^2 d)$ per step (particle setting); for matrix-valued PDEs, block Gram matrices are assembled and solved via Cholesky or iterative solvers.

All operations (kernel evaluations, block operations) are GPU-friendly, and random feature expansions enable acceleration in large-scale contexts (Nguyen et al., 27 Jan 2026).

7. Convergence and Empirical Performance

For continuous A-MWGraD with kernel discretization, convergence rates of $O(1/t^2)$ for geodesically convex and $O(\exp(-\sqrt{\beta} t))$ for strongly geodesically convex objectives are proven, improving on the $O(1/t)$ rate of non-accelerated Wasserstein flows. Empirically, on standard multi-modal and Bayesian learning benchmarks, kernel-based A-MWGraD attains stability, sampling accuracy, and Pareto-optimality more rapidly than baseline methods, confirming theoretical rates and the effectiveness of kernel-based discretization for both scalar and operator-driven evolution tasks (Nguyen et al., 27 Jan 2026). For matrix-valued PDEs, rigorous error and stability bounds confirm robust convergence in practice (Giesl et al., 2017).