A Numerical Study of Combining RBF Interpolation and Finite Differences to Approximate Differential Operators

1. Problem Statement

The paper addresses the challenge of numerically approximating differential operators on scattered (irregular) node layouts, which are common in complex geometries where structured grids are impractical.

• Context: Traditional Finite Difference (FD) methods are efficient but require structured grids. Radial Basis Function-generated Finite Differences (RBF-FD) are a popular meshless alternative that generalizes FD to scattered nodes.
• The Gap: A recent "hybrid" approach was proposed that combines RBF interpolation with traditional FD stencils. While promising, its accuracy and efficiency relative to standard RBF-FD had not been thoroughly benchmarked on standard problems.
• Objective: To compare the accuracy and computational cost of the Hybrid RBF-FD method against the standard RBF-FD method using a 2D Poisson equation as a benchmark.

2. Methodology

A. The Two Approaches Compared

1. Standard RBF-FD:

   • Constructs a local RBF interpolant $\hat{u}$ over a stencil of $n$ nearest neighbors.
   • Applies the differential operator $L$ directly to the interpolant to derive weights $w$ such that $(Lu)(x) \approx \sum w_i u(x_i)$.
   • Requires solving a linear system $Aw = (L\phi)(x)$ for each node.

2. Hybrid RBF-FD (The Proposed Method):

   • Starts with a standard FD stencil (e.g., 5-point or 9-point) defined on a "virtual" grid with spacing $\delta$.
   • Since the virtual points $x + \Delta_i$ often do not exist in the scattered node set, the method uses RBF interpolation to estimate values at these virtual points: $u(x+\Delta_i) \approx \sum w_{ij} u(x_j)$.
   • The final operator approximation combines the FD weights ($a_i$) and the RBF interpolation weights ($w_{ij}$).
   • Key Distinction: This allows the adaptation of existing, high-order FD schemes to meshless settings.

B. Experimental Setup

• Test Problem: 2D Poisson equation ($\Delta u = f$) on a unit square with Dirichlet boundary conditions.
  • Analytical solution: $u(x,y) = \sin(\pi x)\sin(\pi y)$.
• Discretization: Scattered nodes generated with a fill density $h=0.01$.
• Parameters:
  • RBF: Polyharmonic spline of order $k=3$ ($\phi(r) = r^3$).
  • Augmentation: Monomials of degree $m \in \{2, 4, 6, 8\}$.
  • Stencil Size ($n$): Set to $2 \times$ (number of monomials).
  • Virtual Stencil Spacing ($\delta$): Varied via a scaling factor $\sigma$ where $\delta = \sigma \cdot h$.
  • Stencils Tested: Standard 5-point (exact for polynomials up to degree 3) and 9-point (exact up to degree 5).
• Metrics: Maximum and mean relative errors against the analytical solution; computation time for weight generation (Phase 1) and linear system solving (Phase 2).

3. Key Contributions & Findings

A. Accuracy Analysis

• Optimal Parameter ($\sigma \approx 1$): The hybrid method performs significantly better than standard RBF-FD when the virtual stencil spacing $\delta$ matches the node fill density $h$ ($\sigma \approx 1$).
  • For low-order augmentation ($m=2$), the hybrid method reduced relative error by up to one order of magnitude compared to RBF-FD.
  • For $m=4$, the hybrid method with a 9-point stencil outperformed RBF-FD in a specific range of $\sigma$, though the margin was smaller.
• Sensitivity to $\sigma$:
  • Small $\sigma$: Errors diverge due to numerical precision issues (division by small $\delta^2$ in FD formulas).
  • Large $\sigma$: Errors increase because virtual nodes become too distant from the center, degrading the RBF interpolation accuracy.
• Polynomial Reproduction Limits:
  • The hybrid method's accuracy is limited by the lower of the FD stencil's polynomial degree or the RBF augmentation degree $m$.
  • For high augmentation orders ($m=6, 8$), standard RBF-FD eventually outperforms the hybrid method because the FD stencils (max degree 5 for the 9-point) become the bottleneck.
• Alternative Interpolation Strategy: The authors tested using different local stencils for each virtual node (rather than one central stencil). This did not improve accuracy but significantly increased computational cost.

B. Computational Cost

• Phase 1 (Weight Construction):
  • The hybrid method is slower. It requires solving $k$ linear systems (where $k$ is the number of FD stencil points) compared to 1 for RBF-FD.
  • However, since the matrix $A$ is identical for all $k$ systems, the LU decomposition is computed once, and only forward/backward substitutions are repeated.
  • Result: The hybrid method takes roughly 1.3x to 1.5x longer in the setup phase (e.g., 57ms vs 41ms for $m=2, n=12$).
• Phase 2 (System Solving):
  • Both methods assemble a sparse global linear system.
  • Using iterative solvers (BiCGSTAB with ILUT preconditioning), the solution times were nearly identical for both methods, provided $\sigma$ was not extreme.

4. Significance and Conclusion

• Main Insight: The hybrid RBF-FD approach is a viable and often superior alternative to standard RBF-FD for approximating differential operators, provided the virtual stencil spacing is carefully tuned to match the node density.
• Trade-off: The method offers a potential order-of-magnitude gain in accuracy for low-to-moderate polynomial reproduction orders at the cost of a moderate increase in setup time (weight generation).
• Limitations: The advantage diminishes at very high augmentation orders where the fixed polynomial degree of the FD stencil limits the hybrid method's potential.
• Future Work: The authors plan to apply this hybrid approach to more complex problems, specifically hyperbolic PDE systems using staggered grid FD schemes, to leverage the method's ability to incorporate physics-specific stencils into meshless frameworks.

In summary, the paper validates that combining RBF interpolation with traditional FD stencils is a powerful technique that can outperform pure RBF-FD in specific regimes, offering a flexible bridge between structured grid accuracy and meshless flexibility.