Classical signal processing frameworks are typically built around well-behaved functions, such as square-integrable ($L^2$) or infinitely smooth functions. However, real-world phenomena and mathematical abstractions frequently introduce highly singular structures: idealized line impulses, step-like edges, and the ridge-like activation patterns found in deep neural networks.
To analyze these directional features across multiple scales, harmonic analysis relies on geometric transforms like the Radon, $k$-plane, shearlet, and Riesz transforms. To make these tools rigorous when applied to singular structures, a powerful research direction has emerged: extending directional integral transforms to the domain of generalized functions (distributions). By leveraging the twin mechanisms of duality and generalized Fourier slice theorems, this approach allows us to manipulate highly singular objects with the same algebraic ease as classical, smooth functions.
The Core Mechanism: Extension via Duality
How do you compute the transform of an object like a Dirac delta or a line singularity that cannot be evaluated pointwise? The foundational trick is to shift the analytical burden from the rough signal to a space of exceptionally smooth "test functions."
In this duality-based framework, a directional transform $T$ is extended to a distribution $f$ by defining its action through a test function $\phi$. Instead of applying $T$ directly to $f$, we apply its adjoint (or transpose mapping) $T^*$ to $\phi$:
$$\langle Tf, \phi \rangle = \langle f, T^* \phi \rangle$$
Because $\phi$ belongs to a well-behaved test space (such as Schwartz functions or Lizorkin spaces), $T^* \phi$ is smooth and well-defined. This definition ensures that the distributional transform remains perfectly consistent with its classical counterpart.
Classical Domain: f (Smooth Function) ─── T ───> Tf (Transform)
│ │
▼ (Inclusion) ▼ (Inclusion)
Distributional Domain: f (Distribution) ─── T ───> Tf (Defined via T*)
This mechanism provides a robust foundation for directional wavelets. For example, in Continuity properties of the shearlet transform and the shearlet synthesis operator on the Lizorkin type spaces, Pilipović et al. utilized this transpose mapping to extend the shearlet transform to Lizorkin-type distributions. The authors proved that the forward transform and its inverse synthesis operator remain continuous and stable, demonstrating that the reconstruction of a distribution can be represented as an absolutely convergent integral over the standard shearlet group.
Similarly, in Distributional Extension and Invertibility of the k-Plane Transform and Its Dual, Parhi et al. used duality to extend the $k$-plane transform—which generalizes the Radon transform by integrating functions over $k$-dimensional planes in $\mathbb{R}^d$—to the space of tempered distributions.
The Distributional Fourier Slice Theorem
In classical imaging, the Fourier Slice Theorem is a vital computational shortcut. It states that taking a one-dimensional Fourier transform of a Radon projection of a 2D image yields an exact 1D slice of that image's 2D Fourier transform.
When moving to generalized functions, this geometric relationship is not guaranteed to hold without rigorous reformulation. To bridge this gap, Parhi et al. derived a generalized Fourier slice theorem specifically for tempered distributions undergoing the $k$-plane transform.
By proving that this slice identity holds in the sense of distributions, the authors established a rigid link between spatial projections and frequency domains for generalized functions. This theorem explains why distributional transforms work so reliably: they preserve the underlying frequency-domain geometry of the space, allowing researchers to analyze the spectral properties of singular ridges and boundaries without needing pointwise evaluations.
Collapsing Multi-Dimensional Complexity into 1D
Another core intuition of this research direction is that complex, multi-dimensional operators can be simplified by routing them through a directional transform space.
Consider the Riesz transform, which extends the 1D Hilbert transform to multiple dimensions and is essential for isolating directional phase information. In Analysis of the Curvature Tensor from the Viewpoint of Signal Processing, Wietzke et al. demonstrated that a Riesz transform of any order can be completely decomposed into a sequence of 1D Hilbert transforms applied directly within the Radon transform space.
By mapping a multi-dimensional signal into Radon space, the complex multi-dimensional singular integral collapses into a collection of independent 1D operations along projection lines. This interplay simplifies both theoretical analysis and implementation:
- Steerable Wavelets: In Decay Properties of Riesz Transforms and Steerable Wavelets, Ward et al. analyzed how the spatial decay of wavelets is affected by Riesz transforms. By understanding this interaction, the authors established precise conditions (smoothness and vanishing moments) required to construct steerable wavelet frames that decay rapidly in space.
- Fractional Generalizations: To gain finer control over feature extraction, Fu et al. introduced a fractional Riesz transform in Riesz transform associated with the fractional Fourier transform and applications in image edge detection. By acting as a multiplier in the fractional Fourier domain, this operator allows for the simultaneous extraction of amplitude, phase, and orientation, optimizing edge detection across fractional geometric spaces.
Applications: From Inverse Problems to Deep Learning
Building a rigorous distributional theory for directional transforms directly solves practical vulnerabilities in inverse problems and machine learning.
Regularizing Ridge-Like Structures
In imaging inverse problems like computed tomography (CT), we often attempt to reconstruct an object from sparse directional projections. In Distributional Extension and Invertibility of the k-Plane Transform and Its Dual, Parhi et al. leveraged their distributional framework to propose novel regularization techniques. Because the framework natively accommodates distributions, it provides the exact mathematical machinery needed to penalize and recover sparse superpositions of $k$-plane ridges, stabilizing reconstructions from incomplete data.
Explaining the Optimality of Neural Networks
Perhaps the most striking application of this framework lies in deep learning theory. A shallow neural network layer computes a linear combination of ridge functions: $\sum c_i \sigma(\mathbf{w}_i^T \mathbf{x} + b_i)$. These ridge functions are intrinsically linked to the Radon and $k$-plane transforms, which integrate along matching hyperplanes.
In Function-Space Optimality of Neural Architectures with Multivariate Nonlinearities, Unser et al. utilized the $k$-plane transform of distributions to define novel "native" Banach spaces. By applying a regularization operator based on this distributional transform, the authors proved a variational representer theorem. They demonstrated that shallow neural networks with multivariate nonlinearities are not just heuristic choices; they are the mathematically optimal solutions for fitting data within these distribution-derived Banach spaces.
Geometric Signal Representations
By blending directional transforms with flexible linear operations, researchers can capture complex geometric singularities. For example, in An Interplay of Ridgelet and Linear Canonical Transforms, Srivastava et al. formulated the linear canonical ridgelet transform. This method applies a linear canonical wavelet transform to the Radon transform of a signal, offering enhanced geometric control and proving an uncertainty inequality that dictates the fundamental limits of line-singularity localization.
Key Foundational Work
Analysis of the Curvature Tensor from the Viewpoint of Signal Processing (2008) — contributed 3 concept elements to this direction.
Decay Properties of Riesz Transforms and Steerable Wavelets (2013) — contributed 1 concept elements to this direction.