22
7

A general approximation lower bound in LpL^p norm, with applications to feed-forward neural networks

Abstract

We study the fundamental limits to the expressive power of neural networks. Given two sets FF, GG of real-valued functions, we first prove a general lower bound on how well functions in FF can be approximated in Lp(μ)L^p(\mu) norm by functions in GG, for any p1p \geq 1 and any probability measure μ\mu. The lower bound depends on the packing number of FF, the range of FF, and the fat-shattering dimension of GG. We then instantiate this bound to the case where GG corresponds to a piecewise-polynomial feed-forward neural network, and describe in details the application to two sets FF: H{\"o}lder balls and multivariate monotonic functions. Beside matching (known or new) upper bounds up to log factors, our lower bounds shed some light on the similarities or differences between approximation in LpL^p norm or in sup norm, solving an open question by DeVore et al. (2021). Our proof strategy differs from the sup norm case and uses a key probability result of Mendelson (2002).

View on arXiv
Comments on this paper