Optimal approximation using complex-valued neural networks

Authors Paul Geuchen, Felix Voigtlaender

Publication date 29 March 2023

Approximation in the complex plane (30E10) Biharmonic, polyharmonic functions and equations, Poisson's equation in two dimensions (31A30) Approximation by other special function classes (41A30) Multidimensional problems (41A63) Artificial neural networks and deep learning (68T07)

Abstract: We prove a quantitative result for the approximation of functions of regularity

C^{k}

(in the sense of real variables) defined on the complex cube

O m e g a_{n} := [- 1, 1]^{n} + i [- 1, 1]^{n} s u b s e t e q m a t h b b C^{n}

using shallow complex-valued neural networks. Precisely, we consider neural networks with a single hidden layer and

m

neurons, i.e., networks of the form

and show that one can approximate every function in

C^{k} l e f t (O m e g a_{n}; m a t h b b C i g h t)

using a function of that form with error of the order

m^{- k / (2 n)}

as

m o i n f t y

, provided that the activation function

p h i : m a t h b b C o m a t h b b C

is smooth but not polyharmonic on some non-empty open set. Furthermore, we show that the selection of the weights

s i g m a_{j}, b_{j} i n m a t h b b C

and

h o_{j} i n m a t h b b C^{n}

is continuous with respect to

f

and prove that the derived rate of approximation is optimal under this continuity assumption. We also discuss the optimality of the result for a possibly discontinuous choice of the weights.

This page was built for publication: Optimal approximation using complex-valued neural networks

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q6431310)