Abstract:We obtain a new universal approximation theorem for continuous operators on arbitrary Banach spaces using the Leray-Schauder mapping. Moreover, we introduce and study a method for operator learning in Banach spaces $L^p$ of functions with multiple variables, based on orthogonal projections on polynomial bases. We derive a universal approximation result for operators where we learn a linear projection and a finite dimensional mapping under some additional assumptions. For the case of $p=2$, we give some sufficient conditions for the approximation results to hold. This article serves as the theoretical framework for a deep learning methodology whose implementation will be provided in subsequent work.

What problem does this paper attempt to address?

The main problems that this paper attempts to solve are two core issues in operator learning: 1. **Learning operators**: How to learn an operator so that its solution can model data. Specifically, given an operator equation, its solution can well describe or predict some complex phenomena (such as dynamic systems). 2. **Approximation in projection space**: How to approximate the original operator in a finite - dimensional projection space and solve the projected operator equation. That is, find a mapping and a linear projection in the projection space so that these mappings and projections can approximate the original operator with arbitrary precision. To achieve the above goals, the paper proposes a projection method based on orthogonal polynomial bases and proves that under certain conditions (for example, in $L^p$ spaces), this method can achieve universal approximation of operators. Moreover, for the case of $p = 2$ (that is, the Hilbert space $L^2$), the paper gives more specific sufficient conditions to ensure the validity of the approximation results. ### Specific problem summary - **Universal approximation theorem**: The paper proves a new universal approximation theorem through the Leray - Schauder mapping, which is applicable to continuous operators on arbitrary Banach spaces. - **Application of the projection method**: A method based on orthogonal projection is introduced for operator learning in $L^p$ spaces, and the corresponding universal approximation results are derived. - **Approximate solutions to fixed - point problems**: It is studied whether the solutions of the projected equations converge to the solutions of the original operator equations, especially in the framework of fixed - point problems. ### Main contributions 1. **Theoretical framework**: A theoretical framework is provided for deep learning methods, which can approximate operators by learning projections and mappings. 2. **Specific implementation**: Although the specific algorithm implementation will be provided in subsequent work, this paper lays a solid theoretical foundation for future work. 3. **Mathematical tools**: Orthogonal polynomials, Leray - Schauder mappings and other mathematical tools are utilized to ensure the rigor and practicality of the theory. Through these works, the paper provides a new perspective and method for the field of operator learning, especially for those complex systems that are difficult to solve analytically.

Projection Methods for Operator Learning and Universal Approximation

Leray-Schauder Mappings for Operator Learning

Analysis of Approximation by Linear Operators on Variable Lρp(·) Spaces and Applications in Learning Theory

Operator Learning: Algorithms and Analysis

An Approximation Theory for Metric Space-Valued Functions With A View Towards Deep Learning

Operator Learning of Lipschitz Operators: An Information-Theoretic Perspective

Universal Approximation of Operators with Transformers and Neural Integral Operators

Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators

Neural Operator: Learning Maps Between Function Spaces

Nonlocality and Nonlinearity Implies Universality in Operator Learning

Deep Operator Network Approximation Rates for Lipschitz Operators

Learning Theory Viewpoint of Approximation by Positive Linear Operators

Operator learning based on sparse high-dimensional approximation

The $p$-Operator Approximation Property

Approximation on Variable Exponent Spaces by Linear Integral Operators

Learning linear operators: Infinite-dimensional regression as a well-behaved non-compact inverse problem

Universal Approximation of Multiple Nonlinear Operators by Neural Networks

Optimal deep learning of holomorphic operators between Banach spaces

Basis Operator Network: A Neural Network-Based Model for Learning Nonlinear Operators Via Neural Basis

Approximate Representer Theorems in Non-reflexive Banach Spaces

Encoder and Decoder Framework for Operator Approximation