Scattering Networks on Noncommutative Finite Groups

Maria Teresa Arias, Davide Barbieri, Eugenio Hernández · 2025

Read Scattering Networks on Noncommutative Finite Groups by Maria Teresa Arias, Davide Barbieri, Eugenio Hernández (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-20950.txt.

Introduction

Maria Teresa Arias, Davide Barbieri, and Eugenio Hernández carry scattering networks from flat space onto finite groups that need not commute. Their paper from twenty twenty-five grounds the method in wavelets, as in the classic design for reading early neural layers. Group-equivariant networks need such tools on group domains. The authors build wavelets on finite groups and weigh them against the classic ones.

Under set conditions on the coefficients, the transform behaves as users hope. It never stretches distances, holds steady under warps, and keeps total energy. Shifts by group elements on either side pass through cleanly. Deeper layers grow less sensitive to shifts of the signal. Worked cases classify data on abelian and nonabelian groups alike. The paper gives group networks a firm signal-processing base.

Catalog record

CREATORMaria Teresa Arias, Davide Barbieri, Eugenio Hernández
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE85 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-20950

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Scattering Networks on Noncommutative Finite Groups Maria Teresa Arias, Davide Barbieri, Eugenio Hernández Abstract Scattering Networks were initially designed to elucidate the behavior of early layers in Convolutional Neural Networks (CNNs) over Euclidean spaces and are grounded in wavelets. In this work, we introduce a scattering transform on an arbitrary finite group (not necessarily abelian) within the context of group-equivariant convolutional neural networks (G-CNNs). We present wavelets on finite groups and analyze their similarity to classical wavelets. We demonstrate that, under certain conditions in the wavelet coefficients, the scattering transform is non-expansive, stable under deformations, preserves energy, equivariant with respect to left and right group translations, and, as depth increases, the scattering coefficients are less sensitive to group translations of the signal, all desirable properties of convolutional neural networks. Furthermore, we provide examples illustrating the application of the scattering transform to classify data with domains involving abelian and nonabelian groups. Keywords: scattering, machine learning, CNNs, equivariance, groups. 1 Introduction Many interesting problems involve analyzing and manipulating structured data, often real-valued functions defined on domain sets with some structure. This is the case of image recognition, where convolutional neural networks (CNNs) have been extremely successful, giving rise to many generalizations and to a great interest in understanding how they work. The concept of scattering transform was originally introduced in [ 21 ] in L^{2}(\mathbb{R}^{d}) . It can be viewed as a CNN which builds a representation of the data having provable properties. It has a multi-layer structure and uses convolution with wavelets \psi_{\lambda} to propagate signals and convolution with a filter \phi to generate the features (See Figure 1 ). Figure 1 : The structure of the scattering transform of a signal f . In this work, we introduce a scattering transform for signals in L^{2}(G) , where G is a finite group, not necessarily abelian. This generalization is motivated by the need to analyze data exhibiting non-commutative symmetries, which arise naturally in various contexts. For instance, in cryptography, signals defined in the symmetric group S_{n} , representing permutations of data, inherently possess a noncommutative structure. Furthermore, the underlying symmetries of certain data transformations, such as those associated with wavelet or time-frequency analysis, can often be more accurately modeled using affine groups over finite fields. Indeed, extending the scattering transform to this setting is a natural progression, given that convolutional structures are not merely sufficient, but also necessary, for equivariance to the action of a compact group [ 17 ] . To achieve this, we introduce a family of wavelets on L^{2}(G) , adapted to the group structure, and use them to construct the group scattering transform. Our framework is based on representation theory and harmonic analysis, providing a rigorous foundation for the analysis of these signals. We prove that the resulting G -scattering transform exhibits desirable equivariance properties, respecting the underlying group symmetries, and demonstrate its stability, ensuring robustness to small perturbations in the input signal. 1.1 Results Section 2 introduces and motivates our definition of G -wavelets, drawing connections to existing wavelet constructions in the literature. A G -wavelet \psi is defined as a linear combination of the characters \{\chi^{\pi}:\pi\in\widehat{G}\} of the group G , taking into account the degree, d_{\pi} , of the unitary representations \pi that give rise to the character \chi^{\pi}(x)=Tr(\pi(x)),\ x\in G . That is (see Definition 2.2 ) \psi(x)=\sum_{\pi\in\widehat{G}}d_{\pi}\gamma_{\psi}(\pi))\chi^{\pi}(x)\,\quad x\in G\,. The coefficients \{\gamma_{\psi}(\pi):\pi\in\widehat{G}\} will play an important role in proving the result described below. We define the G -wavelet transform of f\in L^{2}(G) as a group convolution with a family of G -wavelets \{\psi_{j}\}_{j=0}^{J}\subset L^{2}(G) , by Wf(j,x)=\frac{1}{|G|}\sum_{h\in G}\psi_{j}(h)f(h^{-1}x)=\psi_{j}\ast f(x)\,. Under the Calderón type condition \mathscr{C}(\pi)=\sum_{j=0}^{J}|\gamma_{\psi_{j}}(\pi)|^{2}=1\,,\quad\forall\ \pi\in\widehat{G}\,, (1) the family of group translations of the G -wavelets form a Parseval frame, i.e. \sum_{j=0}^{J}\|Wf(j,\cdot)\|_{L^{2}(G)}^{2}=\|f\|_{L^{2}(G)}^{2}\,,\ \forall\ f\in L^{2}(G)\,. (2) In fact, conditions ( 1 ) and ( 2 ) are equivalent (see Theorem 2.6 for a more general result). The G -wavelet transform has the desirable property of being equivariant under both left and right group translations. That is, if L_{g}f(x)=f(g^{-1}x) is the left translation operator and R_{g}f(x)=f(xg) is the right translation operator in the group G , then, for all g\in G and all f\in L^{2}(G) , it holds that W(L_{g}f)=L_{g}W(f) and W(R_{g}f)=R_{g}W(f) (see Subsection 4.3 ). Section 3 presents our notion of group scattering transform on finite groups: for a Parseval G -wavelet frame \{\psi_{j}\}_{j=0}^{J}\subset L^{2}(G) , that is, one that satisfies ( 2 ), and denoting by \phi=\psi_{0} , for any m\in\mathbb{N} we define S[\emptyset]f=\phi\ast f\,,\ S[j_{1},\dots,j_{m}]f=\phi\ast|\psi_{j_{m}}\ast\dots|\psi_{j_{1}}\ast f|\dots|. Here m denotes the depth of the convolutional structure. We invite the reader to see Figure 3 that contains a graphical description of this transform. In Section 4 , we prove the key structural properties of this transform. Namely, with Lemma 4.1 we prove non-expansivity, that is \|Sf\|^{2}:=\|\phi\ast f\|_{L^{2}(G)}^{2}+\sum_{m=1}^{\infty}\sum_{j_{1},\dots j_{m}\in\{1,\dots,J\}}\|S[j_{1},\dots,j_{m}]f\|^{2}_{L^{2}(G)}\leq\|f\|^{2}_{L^{2}(G)}\,,\ \forall\ f\in L^{2}(G). This is a nontrivial property that ensures that, despite the exponential growth of the number of coefficients of the scattering transform, the total energy is controlled by the norm of the original signal. In particular, this implies that for all practical purposes, one can restrict to a finite depth of the scattering coefficients since they will become negligibly small as depth increases. We also prove, with Proposition 4.2 , stability with respect to variations on the signals, namely we prove that \|Sf-Sf^{\prime}\|^{2}\leq\|f-f^{\prime}\|^{2}_{L^{2}(G)}\,,\ \forall\ f,f^{\prime}\in L^{2}(G). Note that, since S is not linear, this is not a direct consequence of the non-expansivity. A finer property that we prove, with Theorem 4.6 , is energy preservation, that is \|Sf\|^{2}=\|f\|^{2}_{L^{2}(G)}\,,\ \forall\ f\in L^{2}(G)\,. This holds for the scattering transform of a Parseval frame wavelet if, in addition, we assume that the function \phi=\psi_{0} satisfies the admissibility condition \beta:=\min_{\pi\in\widehat{G}}|\gamma_{\phi}(\pi)|^{2}>0 (see Definition 4.3 ). This same admissibility condition also ensures injectivity of the scattering transform. We will further discuss in Section 5 a relaxed admissibility condition, which relies on a structure property of the unitary irreducible representations of finite groups. Due to the underlying equivariance property of the wavelet construction, we show with Proposition 4.7 that the group scattering transform is equivariant with respect to left and right group translation. Finally, under the previously mentioned admissibility condition, we prove in Theorem 4.8 that as depth increases, an approximate invariance with respect to group translations emerges. More precisely, we show that there exists a positive constant \alpha<1 and a constant C>0 such that \sum_{j_{1},\dots j_{m}\in\{1,\dots,J\}}\|S[j_{1},\dots,j_{m}]L_{g}f-S[j_{1},\dots,j_{m}]f\|^{2}_{L^{2}(G)}\leq C\alpha^{m}\|f\|^{2}_{L^{2}(G)} for all f\in L^{2}(G) . This indicates that the scattering coefficients at increasing depths are less and less sensitive to group translations of the signal, which is a desirable property in convolutional neural networks. Section 6 provides three illustrative examples of applications. In the first one we refine the results of [ 36 ] in the context of the MNIST dataset (Section 6.1 ). After choosing an appropriate wavelet frame, we construct the scattering transform for abelian translations and show accuracy of more than 97% in the classification. In the second example (Section 6.2 ), we demonstrate the utility of the scattering transform for the class of groups \operatorname{Aff}(\mathbb{F}_{p}) in audio classification of signals belonging to two different categories (audio of barks and meows from dogs and cats). Here we first represent audio signals on affine groups by means of time-scale

Read the complete text