Abstract
This article is a continuation of recent research into the off-diagonal-decay properties—often referred to as "localization"—of Moore-Penrose pseudoinverses of bi-infinite matrices. Such properties are critical when matrices arise from the discretization of operators with respect to a frame or basis. While previous work was restricted to specific systems like wavelet or Gabor frames, this paper extends and unifies these results. By establishing localization properties for the general framework of $\alpha$-molecules, the authors provide a simultaneous proof for wavelets, curvelets, shearlets, ridgelets, and hybrid shearlets.
Introduction
In modern signal processing and numerical analysis, functions are often represented using "frames"—sets of building blocks that offer more flexibility than standard orthonormal bases. A desirable property for these frames is localization: the idea that frame elements only interact strongly with their immediate neighbors in space, frequency, and orientation. When an operator is discretized using a localized frame, the resulting matrix is "sparse-ish," meaning most entries far from the diagonal are nearly zero.
However, a significant mathematical challenge arises when we need to reconstruct a signal: inverse closedness. If a matrix $A$ exhibits off-diagonal decay, does its Moore-Penrose pseudoinverse $A^+$ (essential for finding the dual frame) inherit this same decay? Historically, this had to be proven individually for every new system. This paper introduces a unified framework to solve this problem for a broad class of anisotropic systems.
The $\alpha$-Molecule Framework
To move beyond case-by-case proofs, the authors utilize the concept of $\alpha$-molecules. This is a parameterized family of frames where the parameter $\alpha \in [0, 1]$ determines the scaling symmetry of the system:
- $\alpha = 1$: Corresponds to wavelets, which are isotropic (scaling equally in all directions).
- $\alpha = 1/2$: Corresponds to curvelets or shearlets, which follow parabolic scaling.
- $\alpha = 0$: Corresponds to ridgelets.
By treating these diverse systems as points on a single spectrum, the authors can establish "universal" localization results that apply to all of them simultaneously.
Methodology: Abstract Distance and Spectral Gaps
The research employs a two-step strategy to prove localization:
- Abstract Framework: The authors define a generalized "index distance" function $\omega$ that measures the distance between two frame elements in phase space. They prove that for any distance function satisfying a pseudo-triangle inequality, a matrix with off-diagonal decay will have a pseudoinverse with similar decay, provided the matrix possesses a "spectral gap" (meaning it is numerically well-behaved).
- Application to $\alpha$-Molecules: The authors demonstrate that the geometry of $\alpha$-molecules fits this abstract framework. A key technical achievement is proving that the norm of localized matrices is submultiplicative ($||AB|| \le ||A|| \cdot ||B||$), allowing for the use of Neumann series-style arguments to analyze the pseudoinverse.
Key Results
Theorem 2.14: The Main Result
The central contribution of the paper is Theorem 2.14, which establishes that if a matrix $A$ is $N$-localized with respect to an admissible distance $\omega$, its pseudoinverse $A^+$ is $N^+$-localized. This confirms that the dual frames of $\alpha$-molecules remain well-localized, ensuring that reconstruction remains computationally efficient and stable.
Explicit Decay Rates
The paper provides an explicit formula (Equation 2.16) for the decay rate $N^+$ of the pseudoinverse. This formula quantifies how the decay depends on:
- The original decay rate $N$.
- The spectral bounds of the matrix.
- The specific properties of the distance function $\omega$.
Universal Applicability
The authors successfully verify that standard sampling grids for curvelets and shearlets satisfy the necessary conditions for their theorem. This provides a rigorous mathematical foundation for the use of these frames in practical applications like image denoising and compressed sensing.
Discussion and Limitations
While the paper provides a powerful unification, it also highlights certain trade-offs. The decay rate $N^+$ of the pseudoinverse is generally lower than the original decay rate $N$, representing a inherent "decay loss" during the inversion process. Furthermore, the results are highly dependent on the existence of a spectral gap. If a frame is excessively redundant or poorly constructed, the Gram matrix (the matrix of inner products) may become nearly singular, causing the localization of the dual frame to degrade.
Future work may look to extend these "intrinsic" properties from discrete index sets to continuous frame transforms, further broadening the scope of this unified theory of anisotropic localization.
Summary updated