Abstract
This paper demonstrates how composite-exponential-fitting interpolation rules (CEFIR) can be constructed to fit oscillatory functions using not only pointwise values of the function but also its derivatives on a closed and bounded interval. Developed within the framework of exponential-fitting techniques, these rules extend classical composite cubic Hermite interpolating polynomials (CHIP), simplifying to the classical forms as the frequency parameter tends to zero. Comparative examples illustrate the superior performance of CEFIR against traditional interpolation methods for high-frequency data.
Introduction
Interpolating oscillatory data—such as signals, waves, or vibrations—is notoriously difficult for standard polynomial methods. While classical Hermite interpolation, which utilizes both function values and derivatives, is generally accurate for smooth data, it struggles to track high-frequency oscillations unless an impractical number of data points are used.
Existing global "exponential-fitting" methods can handle these oscillations by using trigonometric basis functions instead of polynomials. However, these methods face a major hurdle: numerical instability. As more nodes are added to a global rule to increase accuracy, the condition number of the underlying linear system explodes, rendering the results unreliable. This research introduces CEFIR as a middle ground: a method that is frequency-aware but remains numerically stable through a composite approach.
Methodology
The CEFIR framework is developed through several key steps:
- Exponential Basis: Instead of fitting data to standard polynomials, the rule is designed to be exact for functions of the form $\exp(\pm i\omega x)$, which represent underlying oscillations.
- Two-Node Hermite Rule: The author constructs a basic rule using two nodes, incorporating both the function value $f(x)$ and its derivative $f'(x)$ at each point.
- Composite Strategy: To avoid the instability of global fitting, the rule is applied in a "composite" fashion. The total interval is divided into small subintervals, and the rule is applied piecewise, similar to cubic splines.
- Mathematical Convergence: The method extends classical theory; as the frequency parameter ($\omega$) approaches zero, the CEFIR formulas simplify exactly into the classical composite cubic Hermite interpolating polynomials.
Results and Discussion
The method was evaluated using a synthetic, highly oscillatory test function: $f(x) = x \sin x \cos \omega x + x \cos x \sin \omega x$. The results were compared against CHIP and previously developed global exponential-fitting rules.
As shown in the figure above, the classical CHIP method (dash-dot line) fails to capture the peaks and troughs of the oscillatory function $f(x)$ (solid line). In contrast, CEFIR (dotted line) tracks the oscillatory behavior almost perfectly.
A critical insight of this work is that keeping the interpolation local (composite) rather than global ensures the condition number remains low and constant. While a 7-node global rule can reach a condition number of $3 \times 10^7$, the composite CEFIR rule stays below 25, regardless of how many subintervals are used. This allows CEFIR to incorporate the frequency $\omega$ directly into its coefficients, enabling it to "anticipate" oscillations between data points without sacrificing numerical reliability.
Conclusion
CEFIR provides a robust mathematical tool that works for both smooth, low-frequency data and highly volatile signals. While the current method assumes the frequency $\omega$ is known and requires derivative data ($f'$), it offers a significant improvement over classical polynomials for oscillatory datasets. Future work may focus on integrating frequency estimation techniques and extending the rule to higher-order nodes per subinterval.
Summary updated