Abstract
This paper derives sparse recurrence relations between orthogonal polynomials on a triangle and their partial derivatives, which are analogous to recurrence relations for Jacobi polynomials. We derive these recurrences in a systematic fashion by introducing ladder operators that map an orthogonal polynomial to another by incrementing or decrementing its associated parameters by one. We apply the results to efficiently calculating the Laplacian of polynomial approximations of functions on the triangle, using polynomial degrees in the thousands, i.e., millions of degrees of freedom.
Introduction
Spectral methods are highly valued in numerical analysis for their exponential convergence rates when solving partial differential equations (PDEs). While these methods are well-established for tensor-product domains like rectangles (using Fourier or Chebyshev bases), extending them to triangular domains introduces significant complexity.
Efficiently solving PDEs on a triangle requires a basis of orthogonal polynomials that allows for sparse representations of differential operators. Koornwinder defined a suitable family of polynomials in the 1970s, but the sparse recurrence relations necessary for high-performance computation—specifically those that allow a Laplacian to be represented as a banded matrix—were previously incomplete. Without these relations, spectral methods on triangular meshes become computationally prohibitive at high resolutions.
Methodology: The Ladder Operator Framework
The authors address this gap by developing a systematic framework based on "ladder operators." These operators provide a structured way to navigate the space of orthogonal polynomials by shifting their degrees and parameters.
Univariate Foundation
The derivation begins with 12 ladder operators for univariate Jacobi polynomials. These operators act as fundamental steps, transforming a polynomial of degree $n$ with parameters $(a, b)$ into a state where the degree or parameters are incremented, decremented, or held constant.
Bivariate Extension to the Triangle
This logic is extended to the bivariate family $P_{n,k}^{(a,b,c,d)}(x, y)$ on the triangle, resulting in 24 distinct ladder operators. By composing these operators, the authors derive sparse recurrence relations for three essential operations:
- Differentiation: Expressing partial derivatives in terms of polynomials with shifted parameters.
- Conversion: Mapping between different parameter sets (e.g., from $(a,b,c)$ to $(a+1,b,c)$).
- Multiplication: Handling products of polynomials with the coordinates $x, y$, or $z$.
This structured approach avoids the algebraic complexity of brute-force derivation and ensures that the resulting operator matrices maintain a banded (sparse) structure.
Key Insights and Results
The core innovation of this work is the realization that differential operators, typically dense in a fixed polynomial basis, become sparse when the parameters of the basis are allowed to vary. By dynamically shifting the basis parameters during differentiation, the computational complexity is kept to a minimum.
Performance and Scaling
The proposed method achieves $O(d^2)$ complexity for a polynomial of degree $d$. Given that the total degrees of freedom $N$ is proportional to $d^2$, the method scales at the theoretical limit of $O(N)$. In practical tests using a synthetic function $f(x, y) = \cos(nxy/40)$ on a unit right-angled triangle, the authors successfully computed Laplacians for polynomial degrees in the thousands, representing millions of degrees of freedom.
Numerical Stability
Despite the high polynomial degrees, the relative error remained near machine epsilon ($\sim 10^{-12}$ to $10^{-11}$). This demonstrates the exceptional numerical stability of the derived recurrence relations, which is critical for high-fidelity scientific simulations.
Conclusion and Future Work
This research provides the essential building blocks for high-order spectral methods on triangular domains. While the current focus is on differentiation and Laplacian operators for linear PDEs, the framework is designed for extensibility. Future work will involve applying these sparse relations to solve full linear PDEs on triangles and extending the ladder operator methodology to 3D simplices, such as tetrahedra.
Summary updated