Abstract
This paper presents a systematic framework for constructing explicit, computable basis functions for smooth ($C^m$ conforming) finite element spaces in arbitrary dimensions. By leveraging the dual basis of Bernstein polynomials and a distance-based geometric lattice decomposition, the authors transform the complex problem of enforcing high-order continuity into an efficient, block-lower-triangular computational task. The method is validated through numerical experiments in 2D and 3D, demonstrating optimal convergence rates and practical realizability for complex simulations.
Introduction
Numerical simulations of high-order partial differential equations (PDEs)—such as those governing thin-plate vibrations or complex fluid flows—necessitate "smooth" finite element spaces. These spaces require $C^m$ continuity, meaning functions and their derivatives up to order $m$ must remain continuous across mesh boundaries. While theoretical work has established the existence of these spaces for any dimension and degree of smoothness, the lack of explicit basis functions has historically hindered practical application. Implementation has traditionally been a "dark art," often restricted to manually derived 2D cases. This paper bridges the gap between theory and practice by providing a universal recipe for basis construction.
Methodology
The authors automate the construction of smooth bases by integrating several advanced techniques into a unified framework:
Geometric Lattice Decomposition
The framework utilizes a distance-based structure to organize the lattice points within a simplex (e.g., a triangle or tetrahedron). This allows for a precise association between degrees of freedom (DoFs) and geometric features like vertices, edges, or faces.
Bernstein-Bézier Framework and Dual Bases
Instead of standard polynomials, the method employs the Bernstein basis, favored in computer-aided design for its numerical stability in enforcing smoothness. The core innovation lies in the definition of degrees of freedom. Rather than using difficult-to-compute "integral moments," the authors evaluate functions against the dual basis of Bernstein polynomials. This choice automatically structures the resulting system.
Block Lower Triangular Structure
By strategically selecting these modified DoFs, the relationship between DoFs and basis functions is structured as a block lower triangular matrix. This enables the computation of basis functions via simple forward substitution, avoiding the prohibitive cost of global matrix inversions. The method is implemented in the FEALPy platform, supporting high-performance backends like NumPy, PyTorch, and JAX.
Results and Evaluation
The method was evaluated using standard numerical benchmarks, including trigonometric interpolation and the solution of biharmonic ($m=1$) and triple harmonic ($m=2$) equations on unit squares and cubes.
- Optimal Convergence: Experiments confirmed that interpolation errors decrease at theoretically optimal rates. For instance, $C^1$ elements in 2D using degree $k=7$ polynomials achieved an $L^2$ error reduction of $O(h^8)$.
- 3D Implementation: The authors successfully constructed $C^1$ elements on tetrahedral meshes, a task historically regarded as one of the most challenging in finite element programming.
- Computational Efficiency: The local nature of the basis construction, enabled by the triangular matrix structure, ensures scalability for large-scale industrial simulations.
Conclusion and Future Work
While this framework provides a robust solution for basis construction, challenges remain. For high-smoothness requirements (e.g., $m=2$ in 3D), matrices can become ill-conditioned on coarse meshes. Future research will focus on developing specialized "fast solvers," such as multigrid methods, to handle the large systems generated by these smooth spaces. Additionally, automating the imposition of complex boundary conditions for $C^m$ elements remains a priority for further enhancing the usability of this framework.
Summary updated