The Finite Neuron Method and Convergence Analysis

Jinchao Xu

DOI: https://doi.org/10.4208/cicp.OA-2020-0191

2020-10-04

Abstract:We study a family of $H^m$-conforming piecewise polynomials based on artificial neural network, named as the finite neuron method (FNM), for numerical solution of $2m$-th order partial differential equations in $\mathbb{R}^d$ for any $m,d \geq 1$ and then provide convergence analysis for this method. Given a general domain $\Omega\subset\mathbb R^d$ and a partition $\mathcal T_h$ of $\Omega$, it is still an open problem in general how to construct conforming finite element subspace of $H^m(\Omega)$ that have adequate approximation properties. By using techniques from artificial neural networks, we construct a family of $H^m$-conforming set of functions consisting of piecewise polynomials of degree $k$ for any $k\ge m$ and we further obtain the error estimate when they are applied to solve elliptic boundary value problem of any order in any dimension. For example, the following error estimates between the exact solution $u$ and finite neuron approximation $u_N$ are obtained. $$ \|u-u_N\|_{H^m(\Omega)}=\mathcal O(N^{-{1\over 2}-{1\over d}}). $$ Discussions will also be given on the difference and relationship between the finite neuron method and finite element methods (FEM). For example, for finite neuron method, the underlying finite element grids are not given a priori and the discrete solution can only be obtained by solving a non-linear and non-convex optimization problem. Despite of many desirable theoretical properties of the finite neuron method analyzed in the paper, its practical value is a subject of further investigation since the aforementioned underlying non-linear and non-convex optimization problem can be expensive and challenging to solve. For completeness and also convenience to readers, some basic known results and their proofs are also included in this manuscript.

Numerical Analysis

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is to construct a new class of numerical methods - the Finite Neuron Method (FNM) for solving high - order partial differential equations (PDEs) in arbitrary dimensions. Specifically, the paper studies a family of $H^m$-conforming piecewise polynomial functions based on artificial neural networks. These functions are called the Finite Neuron Method (FNM), and the convergence analysis of this method is provided. Given a general domain $\Omega \subset \mathbb{R}^d$ and a partition $T_h$ on it, how to construct a conforming finite element subspace of $H^m(\Omega)$ with appropriate approximation properties remains a generally unsolved problem. By using the techniques of artificial neural networks, the paper constructs a set of $H^m$-conforming functions consisting of piecewise polynomials of arbitrary degree $k \geq m$, and further obtains the error estimates when they are applied to solve elliptic boundary value problems of arbitrary order and dimension. For example, the following error estimate is obtained: \[ \|u - u_N\|_{H^m(\Omega)} = O(N^{-\frac{1}{2} - \frac{1}{d}}) \] In addition, the paper also discusses the differences and relationships between the Finite Neuron Method and the Finite Element Method (FEM). For example, for the Finite Neuron Method, the underlying finite - element mesh is not given a priori, and the discrete solution can only be obtained by solving a nonlinear and non - convex optimization problem. Although the paper analyzes many desirable theoretical properties of the Finite Neuron Method, its practical value still needs further study because the above - mentioned potential nonlinear and non - convex optimization problem may be both costly and difficult to solve.

The Finite Neuron Method and Convergence Analysis

A neuron-wise subspace correction method for the finite neuron method

An Unconstrained Formulation of Some Constrained Partial Differential Equations and its Application to Finite Neuron Methods

Physics-informed Neural Networks for Functional Differential Equations: Cylindrical Approximation and Its Convergence Guarantees

Neural network-augmented differentiable finite element method for boundary value problems

Neural-Network-Assisted Finite Difference Discretization for Numerical Solution of Partial Differential Equations

Convergence and error control of consistent PINNs for elliptic PDEs

Local Extreme Learning Machines and Domain Decomposition for Solving Linear and Nonlinear Partial Differential Equations

FDM/FEM for nonlinear convection–diffusion–reaction equations with Neumann boundary conditions—Convergence analysis for smooth and nonsmooth solutions

Augmented neural forms with parametric boundary-matching operators for solving ordinary differential equations

Convergence Framework of Deep Learning-based Hybrid Iterative Methods and the Application to Designing a Fourier Neural Solver for Parametric PDEs

Convergence analysis for minimum action methods coupled with a finite difference method

Error analysis for hybrid finite element/neural network discretizations

Global Convergence of Deep Galerkin and PINNs Methods for Solving Partial Differential Equations

Nonconforming Virtual Element Method for $2m$-th Order Partial Differential Equations in $\mathbb R^n$ with $m>n$

Spectral methods for nonlinear functionals and functional differential equations

Subspace method based on neural networks for solving the partial differential equation

Convergence analysis of wide shallow neural operators within the framework of Neural Tangent Kernel

Beyond Finite Layer Neural Networks: Bridging Deep Architectures and Numerical Differential Equations

Error analysis for finite element operator learning methods for solving parametric second-order elliptic PDEs

Can Physics-Informed Neural Networks beat the Finite Element Method?