Abstract
The perturbation of the eigenvalues of a regular Sturm–Liouville problem in normal form which results from a small perturbation of the coefficient function is known to be uniformly bounded. For numerical methods based on approximating the coefficients of the differential equation, this result is used to show that a better bound on the error is obtained when the problem is in normal form. A method having a uniform error bound is presented, and an extension of this method for general Sturm–Liouville problems is proposed and examined.
Introduction
Sturm-Liouville (S-L) problems are essential to mathematical physics, providing the framework for understanding phenomena ranging from the vibration of strings to the free oscillations of the Earth. These problems are typically expressed as second-order differential equations of the form:
$$-(pu')' + qu = \lambda ru$$
While standard numerical techniques—such as Finite Difference methods or the Rayleigh-Ritz method—are highly effective for calculating the first few eigenvalues ($\lambda_k$), they encounter a significant obstacle known as the "High-Frequency Headache." As the index $k$ increases, the eigenfunctions become increasingly oscillatory. Because traditional methods rely on approximating the solution $u(x)$ using polynomials or Taylor series, they struggle to capture these rapid oscillations. Consequently, the error in eigenvalue estimation often grows quadratically or cubically with $k$, rendering high-frequency calculations computationally expensive or inaccurate.
Methodology
This paper proposes a shift in numerical strategy: rather than approximating the rapidly oscillating solution, we should approximate the differential equation itself. The approach consists of three primary stages:
1. Liouville Normal Form Transformation
The general S-L problem is first transformed into its "normal form":
$$-u'' + su = \lambda u$$
This transformation is a critical preprocessing step. In this form, the eigenvalues are significantly less sensitive to small perturbations in the coefficient function $s(x)$ than they are in the general form. This mathematical "cleanup" provides the stability necessary for uniform error bounds.
2. Piecewise Constant Approximation
The coefficient function $s(x)$ is approximated by a piecewise constant function $\tilde{s}(x)$, typically using midpoint interpolation. While the eigenfunction $u(x)$ wiggles violently at high frequencies, the coefficient function $s(x)$ is usually smooth and well-behaved. Approximating the "rules" of the physics (the coefficients) is inherently more stable than approximating the "result" (the solution).
3. The Hybrid Strategy and Matrix Implementation
While coefficient approximation is superior to standard methods, it is not perfectly uniform across the entire spectrum. To achieve true uniformity, the authors propose a hybrid approach:
- Numerical Phase: Use the piecewise constant approximation for the lower half of the spectrum.
- Asymptotic Phase: Switch to a simple asymptotic expansion for higher-order eigenvalues.
This hybrid method is implemented via a tridiagonal matrix system, which can be solved efficiently using standard linear algebra techniques such as Gaussian elimination and secant methods.
Key Insights and Theoretical Results
The central contribution of this research is the formal proof of uniform stability. In Theorem 2.1, the authors demonstrate that when a problem is in Liouville normal form, the error in the eigenvalue is bounded by the $L_2$ norm of the difference between the true and approximated coefficient functions:
$$|\lambda_k - \tilde{\lambda}_k| \leq ||s - \tilde{s}||_2$$
Crucially, this bound is independent of $k$. This means that a small change in the potential $s$ results in a bounded change in the eigenvalue, regardless of how high the frequency is. By combining numerical methods (which excel at low frequencies) with asymptotic formulas (which become increasingly accurate at high frequencies), the authors achieve a uniform $O(h^2)$ error bound across the entire spectrum.
Numerical Evaluation
The method was tested on two synthetic benchmark problems:
- Problem I: A standard S-L problem in normal form on the interval $(0, 1)$.
- Problem II: A complex general S-L problem on the interval $(1, e)$, requiring a Liouville transformation.
In tests using only 16 intervals ($N=16$), the hybrid method maintained a consistent error level for the first 40 eigenvalues. In contrast, standard methods would have seen the error grow by orders of magnitude over the same range. This demonstrates that high-index eigenvalues can be obtained with significantly less computational effort than refining a standard finite-difference mesh to an extreme degree.
Limitations and Future Work
Despite its advantages, the method has specific requirements. It relies on the ability to perform a Liouville transformation; if the original coefficients ($p, q, r$) are noisy or only known at discrete points, the transformation itself may introduce errors. Additionally, the matrix approach can occasionally produce "extraneous zeros"—mathematical artifacts that must be identified and discarded. Finally, the $O(h^2)$ error bound assumes that the coefficient function $s(x)$ is at least twice continuously differentiable ($C^2$). Future research is needed to evaluate performance on non-smooth or singular problems.
Summary updated