Abstract:In this paper, we develop a generalized theory of convolutional signal processing and neural networks for Reproducing Kernel Hilbert Spaces (RKHS). Leveraging the theory of algebraic signal processing (ASP), we show that any RKHS allows the formal definition of multiple algebraic convolutional models. We show that any RKHS induces algebras whose elements determine convolutional operators acting on RKHS elements. This approach allows us to achieve scalable filtering and learning as a byproduct of the convolutional model, and simultaneously take advantage of the well-known benefits of processing information in an RKHS. To emphasize the generality and usefulness of our approach, we show how algebraic RKHS can be used to define convolutional signal models on groups, graphons, and traditional Euclidean signal spaces. Furthermore, using algebraic RKHS models, we build convolutional networks, formally defining the notion of pointwise nonlinearities and deriving explicit expressions for the training. Such derivations are obtained in terms of the algebraic representation of the RKHS. We present a set of numerical experiments on real data in which wireless coverage is predicted from measurements captured by unmaned aerial vehicles. This particular real-life scenario emphasizes the benefits of the convolutional RKHS models in neural networks compared to fully connected and standard convolutional operators.

What problem does this paper attempt to address?

### What problem does this paper attempt to solve? This paper aims to develop a generalized theory for convolutional signal processing and neural networks in Reproducing Kernel Hilbert Spaces (RKHS). Specifically, through the Algebraic Signal Processing (ASP) theory, the author shows how to formally define multiple algebraic convolution models in RKHS and proves that any RKHS can induce an algebraic structure, and the elements of these algebraic structures determine the convolution operators acting on the elements of RKHS. The main objectives include: 1. **Achieve scalable filtering and learning**: Through the convolution model as a by - product, while taking advantage of the known benefits of processing information in RKHS. 2. **Expand the application range of convolutional signal models**: Show how to use algebraic RKHS to define convolutional signal models on groups, graph theory, and traditional Euclidean signal spaces. 3. **Construct convolutional networks**: Based on the algebraic RKHS model, formally define point - wise nonlinearity and derive explicit expressions for training. ### Main contributions of the paper 1. **General convolutional signal model for one - dimensional RKHS signals**: - The author derives a general convolutional signal model suitable for one - dimensional RKHS signals and proves that the classical convolution operation is a special case of this model (when the sinc function is chosen as the kernel). This provides a generalization of standard processing techniques and simplifies the extension of the convolution operator to a wider domain. 2. **General Algebraic Signal Model (ASM) for any RKHS**: - For any RKHS with a domain of monomer or group structure, the author derives an ASM and shows that the representational ability of RKHS is transferred to the algebra through the multiplication operation. This enables convolutional signal processing on groups without using Haar measure integration, thereby reducing complexity and computational cost. 3. **Universal Convolutional Neural Network in RKHS space**: - The author introduces a universal Convolutional Neural Network (CNN) suitable for RKHS space, where the convolution operator naturally arises from the structural properties of RKHS. They also specify a family of nonlinear operators mapping RKHS to itself, ensuring continuity with respect to the norm induced by the Hilbert space structure. In addition, they derive the expression of the steepest - descent method for training the architecture. ### Experimental verification To verify the above theory, the author conducted a series of numerical experiments, especially using real - data collected by drones to predict wireless coverage. These experiments show that the RKHS - based convolution model solves practical problems in neural network architectures and plays a central role in the signal representation involved in physical descriptions. ### Summary This paper solves the problem of efficient convolution operations in RKHS by introducing generalized convolutional signal models and algebraic signal processing methods, and shows its application potential in convolutional neural networks. This not only provides new theoretical tools for the signal processing field but also shows significant advantages in practical applications.

Convolutional Filtering with RKHS Algebras

Learning in RKHM: a $C^*$-Algebraic Twist for Kernel Machines

An Alternate View on Optimal Filtering in an RKHS

An algebra structure for reproducing kernel Hilbert spaces

Lie Group Algebra Convolutional Filters

Non Commutative Convolutional Signal Models in Neural Networks: Stability to Small Deformations

The Generalized Complex Kernel Affine Projection Algorithms

Algebraic Neural Networks: Stability to Deformations

Sparse multiresolution representations with adaptive kernels

Nonuniform sampling, reproducing kernels, and the associated Hilbert spaces

Tensor Decomposition Meets RKHS: Efficient Algorithms for Smooth and Misaligned Data

Generalization Properties of Hyper-Rkhs and Its Applications

RKFNet: A Novel Neural Network Aided Robust Kalman Filter

Finite Dimensional Filters with Non-Linear Drift IX Construction of Finite Dimensional Estimation Algebras of Non-Maximal Rank

Training Neural Networks Using Reproducing Kernel Space Interpolation and Model Reduction

Convolutional networks and applications in vision

Non-linear Convolution Filters for CNN-based Learning

Deep Learning with Kernels through RKHM and the Perron-Frobenius Operator

Kernel Operator-Theoretic Bayesian Filter for Nonlinear Dynamical Systems

An Analytic Solution for Kernel Adaptive Filtering