Abstract
We analyze the representation of periodic signals in a scaling function basis. This representation is sufficiently general to include widely used approximation schemes like wavelets, splines, and Fourier series representations. We derive a closed-form expression for the approximation error in the scaling function representation. The error formula takes the simple form of a Parseval-like sum, weighted by an appropriate error kernel. This formula may be useful in choosing the right representation for a class of signals. We also experimentally verify the theory in the particular case of the description of closed curves.
Introduction
Classical sampling theory is primarily developed for signals in the $L_2(\mathbb{R})$ space—signals that are square-integrable over the entire real line. However, many practical applications involve periodic signals, such as the closed contours of objects in computer vision or repeating physiological signals like EKGs.
Applying standard $L_2(\mathbb{R})$ error analysis to periodic signals is mathematically problematic. Because periodic signals repeat infinitely, their Fourier transforms consist of discrete Dirac impulses, rendering standard integral-based error measures undefined. While researchers have previously used "periodized" basis functions as a workaround, a general, predictive error formula that allows for direct comparison between different representation schemes (e.g., splines vs. wavelets vs. sinc) has been lacking.
Theoretical Framework
The authors propose a general predictive error formula based on the Fourier coefficients of the input signal. The derivation relies on three key pillars:
- Periodization: Adapting standard synthesis and analysis equations using $N$-periodized versions of the generating functions.
- Shift-Invariance: To account for phase sensitivity (how error changes based on sample placement), the authors define a mean square error averaged over all possible time shifts.
- The Error Kernel ($E(\omega)$): The total error is expressed as a "Parseval-like" sum: $$\eta_s(N) = \sqrt{\sum |S(k)|^2 E(2\pi k / N)}$$ Here, $S(k)$ represents the signal's harmonics, and $E(\omega)$ is the error kernel specific to the chosen sampling method. This kernel decomposes into a minimum possible error ($E_{min}$) for optimal orthogonal projection and a residual error ($E_{res}$) for non-optimal interpolation.
Experimental Validation
To validate the theoretical model, the authors utilized a complex real-world dataset: a polygonal representation of the map of Switzerland consisting of 807 edges. The $(x(t), y(t))$ coordinates of the boundary were treated as two separate periodic signals.
Quantitative Results
The experimental results closely mirrored the theoretical predictions. Notably, the analysis revealed that cubic splines consistently outperform Sinc interpolation (the standard Fourier approach) at lower sampling rates.
As shown in the figure above, cubic splines provide a performance gain of approximately 1 dB over Sinc interpolation across a wide range of sampling densities. This quantitative advantage is critical for applications requiring high precision with minimal data storage.
Qualitative Comparison
The difference in performance is even more striking when viewed qualitatively. Sinc interpolation is a "global" method; a sharp feature in one part of the signal can cause "ringing" or "looping" artifacts (the Gibbs phenomenon) throughout the reconstruction. In contrast, splines are "local," making them more robust for representing shapes with sharp transitions and intricate details.
When resampling the Switzerland map at 45 points, Sinc interpolation creates unnatural loops and artifacts at the borders, whereas the cubic spline representation remains much more faithful to the original geometry.
Conclusion
This research provides a practical tool for engineers to predict approximation quality and select the most efficient sampling scheme for periodic data. By demonstrating that cubic splines can outperform the "gold standard" Sinc interpolation in specific regimes, the paper highlights the importance of choosing basis functions that match the local characteristics of the signal. Future work may address limitations such as the assumption of uniform sampling and the extension of this analysis to multi-dimensional periodic data.
Summary updated