1: \begin{abstract}

2:   We study a family of $H^m$-conforming piecewise polynomials based on

3:   artificial neural network, named as the {\it finite neuron method}

4:   (FNM), 

5:   for numerical solution of $2m$-th order partial differential

6:   equations in $\mathbb{R}^d$ for any $m,d \geq 1$ and then provide

7:   convergence analysis for this method.  Given a general domain

8:   $\Omega\subset\mathbb R^d$ and a partition $\mathcal

9:   T_h$ of $\Omega$, it is still an open problem in general how to

10:   construct conforming finite element subspace of $H^m(\Omega)$ that

11:   have adequate approximation properties. By using techniques from

12:   artificial neural networks, we construct a family of

13:   $H^m$-conforming set of functions consisting of piecewise

14:   polynomials of degree $k$ for any $k\ge m$ and we further obtain the

15:   error estimate when they are applied to solve elliptic boundary

16:   value problem of any order in any dimension.  For example, the

17:   following error estimates between the exact solution $u$ and finite

18:   neuron approximation $u_N$ are obtained. 

19: $$

20: \|u-u_N\|_{H^m(\Omega)}=\mathcal O(N^{-{1\over 2}-{1\over d}}).

21: $$

22: Discussions will also be given on the difference and relationship

23: between the finite neuron method and finite element methods (FEM).  For

24: example, for finite neuron method, the underlying finite element grids

25: are not given a priori and the discrete solution can only be obtained

26: by solving a non-linear and non-convex optimization problem.  

27: %

28: Despite of  many desirable theoretical properties of the finite

29: neuron method analyzed in the paper, its practical value is a subject

30: of further investigation since the aforementioned underlying

31: non-linear and non-convex optimization problem can be expensive and

32: challenging to solve.  For completeness and also convenience to

33: readers, some basic known results and their proofs are also included

34: in this manuscript.

35: \end{abstract}

36: