[Paper Review] Classification of Generalized Multiresolution Analyses

Lawrence W. Baggett, Veronika Furst+2 more|ArXiv.org|Oct 28, 2009

Mathematical Analysis and Transform Methods7 references3 citations

TL;DR

This paper introduces a classification framework for generalized multiresolution analyses (GMRAs) in abstract Hilbert spaces using two key parameters: a multiplicity function $m$ and a matrix-valued filter function $H$. It establishes conditions under which these parameters uniquely determine a GMRA, provides a constructive procedure to build a canonical GMRA from any valid $(m, H)$ pair, and defines an equivalence relation on GMRAs based on their $(m, H)$ invariants, unifying classical and abstract wavelet constructions including fractal and direct limit examples.

ABSTRACT

We discuss how generalized multiresolution analyses (GMRAs), both classical and those defined on abstract Hilbert spaces, can be classified by their multiplicity functions $m$ and matrix-valued filter functions $H$. Given a natural number valued function $m$ and a system of functions encoded in a matrix $H$ satisfying certain conditions, a construction procedure is described that produces an abstract GMRA with multiplicity function $m $ and filter system $H$. An equivalence relation on GMRAs is defined and described in terms of their associated pairs $(m,H)$. This classification system is applied to classical examples in $L^2 (\mathbb R^d)$ as well as to previously studied abstract examples.

Motivation & Objective

• To unify diverse wavelet constructions—classical, fractal, direct limit, and tensor product examples—under a single classification framework for generalized multiresolution analyses (GMRAs).
• To identify the minimal set of invariants that uniquely characterize a GMRA in an abstract Hilbert space equipped with dilation and translation operators.
• To develop a constructive procedure that generates a canonical GMRA from any given pair $(m, H)$ satisfying specific analytic and algebraic conditions.
• To define and characterize an equivalence relation on GMRAs in terms of their associated multiplicity functions $m$ and filter matrices $H$.

Proposed method

• The classification relies on Stone’s Theorem to decompose the translation representation on $V_0$ via a multiplicity function $m$ on the dual group $\widehat{\Gamma}$, which encodes the spectral multiplicity of characters.
• The filter matrix $H$ encodes the action of the dilation operator on the Fourier side, generalizing classical scalar filters to matrix-valued functions that describe how $V_1$ relates to $V_0$ via inverse dilation.
• A construction procedure is developed that starts from a finite, a.e. defined multiplicity function $m$ and a matrix-valued function $H$ satisfying certain isometric and consistency conditions, producing a Hilbert space with a nested sequence of subspaces satisfying GMRA axioms.
• The equivalence of GMRAs is defined via unitary equivalence that preserves both the multiplicity function $m$ and the filter matrix $H$, leading to a complete invariant system.
• The framework is applied to classical $L^2(\mathbb{R}^d)$ examples and abstract constructions, including direct limits and tensor products of GMRAs.
• The paper uses the Haar GMRA and its tensor product with a dyadic group-based GMRA to construct new examples with non-trivial intersection, which are corrected via tensoring to ensure trivial intersection in the limit.

Experimental results

Research questions

• RQ1 How can generalized multiresolution analyses (GMRAs) in abstract Hilbert spaces be classified using intrinsic invariants?
• RQ2 What conditions on a multiplicity function $m$ and a matrix-valued filter $H$ guarantee the existence of a GMRA with those parameters?
• RQ3 How can a canonical GMRA be explicitly constructed from a given pair $(m, H)$ satisfying the required conditions?
• RQ4 In what sense are two GMRAs equivalent, and how can this equivalence be characterized in terms of their associated $(m, H)$ pairs?
• RQ5 Can the classification framework unify known examples such as the Journe wavelet, Cohen wavelet, and wavelets on fractals?

Key findings

• The paper establishes that a GMRA is completely classified by its multiplicity function $m$ and matrix-valued filter function $H$, with equivalence of GMRAs corresponding exactly to equivalence of their $(m, H)$ pairs.
• A constructive procedure is provided that generates a canonical GMRA from any finite, a.e. defined multiplicity function $m$ and matrix-valued filter $H$ satisfying the required isometric and consistency conditions.
• The example of the Journe wavelet is shown to be inequivalent to the Haar wavelet due to differing filter functions $H$, with the latter having $|h(0)| = \sqrt{2}$, which forces the scaling function to vanish under iteration.
• The construction of a GMRA on the infinite product group $\Gamma = \bigoplus_{i=-\infty}^{\infty} \mathbb{Z}_2$ fails due to non-trivial intersection $\cap V_j$, but this is resolved by tensoring with the standard Haar GMRA on $L^2(\mathbb{R})$, yielding a valid GMRA with $m \equiv 1$ and $\widetilde{m} \equiv 3$.
• The paper constructs two inequivalent GMRAs using the same $m$ and $\widetilde{m}$ but different filters: one with $h_1 = \sqrt{2}\chi_{\{1\}_0 \times \prod_{i=1}^\infty \mathbb{Z}_2_i}$ and another with $h_1' = \chi_{\{1\}_0 \times \cdots} - \chi_{\{-1\}_0 \times \cdots}$, demonstrating distinct filter behavior.
• The tensor product construction with the Haar GMRA yields new fractal-like wavelets, where the filters $h_1'$ and $g_1'$ introduce non-trivial phase behavior not present in classical examples.