20
5

Noncompact uniform universal approximation

Abstract

The universal approximation theorem is generalised to uniform convergence on the (noncompact) input space Rn\mathbb{R}^n. All continuous functions that vanish at infinity can be uniformly approximated by neural networks with one hidden layer, for all activation functions φ\varphi that are continuous, nonpolynomial, and asymptotically polynomial at ±\pm\infty. When φ\varphi is moreover bounded, we exactly determine which functions can be uniformly approximated by neural networks, with the following unexpected results. Let Nφl(Rn)\overline{\mathcal{N}_\varphi^l(\mathbb{R}^n)} denote the vector space of functions that are uniformly approximable by neural networks with ll hidden layers and nn inputs. For all nn and all l2l\geq2, Nφl(Rn)\overline{\mathcal{N}_\varphi^l(\mathbb{R}^n)} turns out to be an algebra under the pointwise product. If the left limit of φ\varphi differs from its right limit (for instance, when φ\varphi is sigmoidal) the algebra Nφl(Rn)\overline{\mathcal{N}_\varphi^l(\mathbb{R}^n)} (l2l\geq2) is independent of φ\varphi and ll, and equals the closed span of products of sigmoids composed with one-dimensional projections. If the left limit of φ\varphi equals its right limit, Nφl(Rn)\overline{\mathcal{N}_\varphi^l(\mathbb{R}^n)} (l1l\geq1) equals the (real part of the) commutative resolvent algebra, a C*-algebra which is used in mathematical approaches to quantum theory. In the latter case, the algebra is independent of l1l\geq1, whereas in the former case Nφ2(Rn)\overline{\mathcal{N}_\varphi^2(\mathbb{R}^n)} is strictly bigger than Nφ1(Rn)\overline{\mathcal{N}_\varphi^1(\mathbb{R}^n)}.

View on arXiv
Comments on this paper