Harmonic spline series representation of scaling functions

Thierry Blu, Michaël UnserProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2003 · DOI

TL;DR

This paper introduces a novel way to represent any compactly supported scaling function as a sum of "harmonic splines." This breakthrough allows for the exact evaluation of scaling functions at any point, including irrational values, which was previously a significant challenge. The work also uncovers new types of scaling functions, the harmonic splines themselves, with potential implications for signal processing and approximation theory.

What Problem Does This Paper Address?

Scaling functions are fundamental building blocks in wavelet analysis, enabling the decomposition of signals into different frequency components. A key property of these functions is their "two-scale relation," which mathematically defines how a function at one scale relates to itself at a finer scale. While this relation is crucial, it makes it notoriously difficult to express scaling functions in a closed-form, explicit manner, especially for arbitrary points. For most standard scaling functions, like Daubechies' wavelets, we can only compute their values exactly at rational numbers. Evaluating them at irrational numbers (like $\pi$) is generally impossible. This paper tackles this limitation by providing an explicit, time-domain representation for any compactly supported scaling function, bridging the gap between theoretical properties and practical evaluation.

Methods

The authors develop a two-step approach. First, they "un-localize" the compactly supported scaling function $\phi(x)$ into a non-compactly supported, self-similar function $\rho(x)$, which they call a "central basis function." This is achieved by expressing $\phi(x)$ as a filtered version of $\rho(x)$. The second, and most significant, step involves decomposing this central basis function $\rho(x)$ into a sum of "harmonic splines." This decomposition is achieved by leveraging the 1-periodicity of a related auxiliary function $u(x)$ and expressing it as a Fourier series. The resulting harmonic splines are characterized by complex degrees, which, when summed appropriately, reconstruct the original scaling function. The paper also introduces a generalized finite difference filter to localize these harmonic splines.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Universal Representation: The core insight is that any compactly supported scaling function can be represented as a sum of harmonic splines. This is a powerful generalization beyond existing methods.
  • Exact Evaluation at Any Point: The harmonic spline decomposition allows for the exact evaluation of scaling functions at any point, including irrational numbers. This is a significant practical advancement.
  • Discovery of Harmonic Splines: The paper introduces "harmonic splines" as a new class of functions. These are not typically compactly supported but are crucial components in the decomposition.
  • Fractal Structure: The approximation of scaling functions using a finite number of harmonic splines reveals a fractal structure, which becomes more apparent with more terms in the sum.
  • Connection to Regularity: The real part of the complex degree of the harmonic splines is shown to coincide with the Hölder regularity of Daubechies scaling functions, suggesting a deep connection between the decomposition and the smoothness properties of the original functions.

Key Results

The main result is Theorem 3.1, which states that every compactly supported scaling function $\phi(x)$ can be expressed as:

$\phi(x) = \sum_{k \ge 0} \sum_{n \in \mathbb{Z}} \gamma_{k,n} \beta_{\alpha + \frac{2i\pi n}{\log 2}}(x-k)$

where $\beta$ are the harmonic splines and $\gamma_{k,n}$ are specific coefficients derived from the original scaling function's properties. The paper demonstrates this by showing an approximation of a Daubechies scaling function of length 4 using a finite number of terms in the harmonic spline expansion, illustrating the convergence towards the original function as more terms are included.

Limitations and Future Work

The authors acknowledge that the harmonic splines themselves are usually not compactly supported, which is a departure from traditional splines. They also note that truncating the harmonic spline expansion (i.e., using a finite number of terms) results in an approximation that satisfies the two-scale difference equation but is not necessarily in $L^1$ because its Fourier transform is not continuous at $\omega = 0$. An open question remains whether it is always possible to find a "localization filter" such that the resulting function is in $L^2$. The paper suggests that the filter defined by a specific integral involving $\rho(x)$ is a good candidate for this goal. Future work could explore the properties and applications of these newly discovered harmonic splines in various signal processing and approximation tasks.

Summary updated

Abstract

We present here an explicit time-domain representation of any compactly supported dyadic scaling function as a sum of harmonic splines. The leading term in the decomposition corresponds to the fractional splines that have recently been defined by the authors as a continuous-order generalization of the polynomial splines.
BibTeX
@inproceedings{blu2003harmonic,
  author = {Thierry Blu and Michaël Unser},
  title = {{Harmonic spline series representation of scaling functions}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {2003},
  doi = {10.1117/12.507293},
}

Loading related content...

Figure