<title>Oblique projections in discrete signal subspaces of l<formula><inf><roman>2</roman></inf></formula> and the wavelet transform</title>

Akram Aldroubi, Michaël UnserProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 1994 · DOI

TL;DR

This paper establishes a direct link between oblique projections in discrete signal spaces and perfect reconstruction filter banks, providing a discrete-time analogue to existing continuous-time wavelet theory. This work offers a unified framework for understanding biorthogonal wavelet decompositions in digital signal processing.

What Problem Does This Paper Address?

The field of wavelet transforms has seen significant advancements, particularly in understanding decompositions like orthogonal, semi-orthogonal, and biorthogonal wavelets. While continuous-time theories, like those by Cohen, Daubechies, and Feauveau, elegantly interpret these as oblique projections in function spaces (L²), a purely discrete-time perspective for digital signal processing has been less developed. Specifically, while the connection between perfect reconstruction filter banks and discrete biorthogonal wavelets is known, the underlying projection operators and their corresponding approximation spaces haven't been explicitly characterized in a way that directly parallels the continuous case. This paper aims to bridge this gap by investigating discrete multiresolutions and wavelets through the lens of oblique projections in the discrete sequence space (l²).

Methods

The authors generalize the concept of multiresolution analysis from continuous function spaces (L²) to discrete sequence spaces (l²). They introduce the notion of "shift-invariant sequence spaces" generated by a single sequence and its translations. The core of their method lies in defining and analyzing "oblique projections" in these discrete spaces. An oblique projection $P_{S|V}$ projects a signal onto a subspace $S$ in a direction orthogonal to another subspace $V$. They show that for specific classes of discrete multiresolutions and their associated wavelet spaces, these oblique projections can be realized using perfect reconstruction filter banks. This establishes a strong equivalence between the abstract concept of oblique projections and the practical implementation via filter banks. The paper also delves into the mathematical conditions for the existence and uniqueness of these projections, relating them to the invertibility of certain operators derived from signal correlations.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Discrete Analogue of Continuous Theory: The paper successfully translates the continuous-time oblique projection framework for wavelets into the discrete-time domain, providing a unified theoretical basis for digital signal processing applications.
  • Oblique Projections as the Unifying Concept: Oblique projections are presented as the fundamental concept that links shift-invariant spaces, multiresolutions, and perfect reconstruction filter banks in the discrete setting.
  • Filter Banks Implement Projections: A key finding is that the decomposition and reconstruction algorithms for certain discrete multiresolutions and their wavelet spaces can be directly implemented using perfect reconstruction filter banks. This means that the abstract mathematical concept of an oblique projection has a concrete, implementable counterpart.
  • Error Bounds and Approximation Quality: The paper provides error bounds for these oblique projections, quantifying how well a signal can be approximated. This is crucial for understanding the performance of wavelet-based compression and denoising algorithms.
  • Generalization of Biorthogonal Wavelets: The work offers a discrete-time analogue to the Cohen-Daubechies-Feauveau results on biorthogonal wavelets, extending their insights to the digital realm.

Key Results

  • The paper defines discrete shift-invariant spaces and establishes conditions (Theorem 1) for a sequence to form a Riesz basis for such a space, which is essential for well-defined projections.
  • It characterizes the oblique projection operator $P_{S|V}$ in discrete shift-invariant spaces and provides conditions for its existence and uniqueness (Theorem 2), linking it to the invertibility of a down-sampled correlation operator.
  • The "spectral coherence" ($p_{S,V}$) is introduced as a measure of the angle between discrete shift-invariant spaces, which dictates the well-definedness and quality of the oblique projection.
  • Theorem 4 provides an error bound for the oblique projection, showing that the approximation error is related to the spectral coherence.
  • The paper defines discrete multiresolutions and homogeneous discrete multiresolutions in l², and establishes their connection to filter banks.
  • Theorem 9 demonstrates a recursive relationship for computing oblique projections in nested multiresolution spaces, leading to a decomposition algorithm (Figure 1 and Algorithm (31)).
  • Theorem 11 shows that the error in a biorthogonal wavelet decomposition can be viewed as an oblique projection onto a wavelet space, further solidifying the connection between projections and filter banks.

Limitations and Future Work

  • Focus on Homogeneous Multiresolutions: While the paper discusses general discrete multiresolutions, many of the specific results and theorems (e.g., Theorem 8, 9, 10, 11) are derived for homogeneous multiresolutions. The extension of these results to non-homogeneous cases might be a direction for future work.
  • Theoretical Framework: The paper primarily focuses on the theoretical underpinnings of oblique projections and their relationship to filter banks. Practical implementation details and performance evaluations on specific signal processing tasks (e.g., image compression, denoising) are not extensively covered.
  • Generalization of "Angle": While the spectral coherence $p_{S,V}$ is defined, its direct interpretation and relationship to other geometric notions of "angle" in discrete spaces could be further explored.
  • Comparison with Existing Discrete Theories: The paper mentions Rioul's work on discrete multiresolutions but suggests their theory doesn't explicitly formulate oblique projections. Further comparison and integration with such existing discrete-time theories could be beneficial.

Summary updated

Abstract

We study the general problem of oblique projections in discrete shift-invariant spaces of l<SUB>2</SUB> and we give error bounds on the approximation. We define the concept of discrete multiresolutions and wavelet spaces and show that the oblique projections on certain subclasses of discrete multiresolutions and their associated wavelet spaces can be obtained using perfect reconstruction filter banks. Therefore we obtain a discrete analog of the Cohen-Daubechies- Feauveau results on biorthogonal wavelets.
BibTeX
@inproceedings{aldroubi1994lt,
  author = {Akram Aldroubi and Michaël Unser},
  title = {{\&lt;title\&gt;Oblique projections in discrete signal subspaces of l\&lt;formula\&gt;\&lt;inf\&gt;\&lt;roman\&gt;2\&lt;/roman\&gt;\&lt;/inf\&gt;\&lt;/formula\&gt; and the wavelet transform\&lt;/title\&gt;}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {1994},
  doi = {10.1117/12.188795},
}

Loading related content...

Figure