no. 16 · Tuesday 17 November 2026
Second scaling regime in high-dimensional function approximation and curse of dimensionality
Abstract
High-dimensional function approximation is often subject to the curse of dimensionality, commonly described by a hypercube scaling regime with a large expense of $R^d$ grid points, where $R$ denotes the side length of a hypercube in $d$ dimensions. Here we show that a second scaling regime, which we call the lattice ball scaling regime, also arises and is relevant for practical applications. In this regime, the expense scales as $d^{R^2}/(R^2!)$ and therefore grows polynomially with $d$. This regime arises without additional special assumptions of special types of function spaces or error norms. The lattice ball scaling regime occurs when $R^2\ll d$, due to the discrete nature of counting the expense as the number of lattice points inside a hypersphere. Here we investigate this regime, which arises from the classical problem of approximating high-dimensional functions from common spaces of analytic or $m$-times continuously differentiable functions using orthogonal series (spectral method) such as Fourier, Chebyshev, and Legendre polynomial series. For many applications, the expense in the lattice ball scaling regime can be substantially smaller than that of the hypercube scaling regime, making accurate approximations more accessible. Numerical illustrations include classifying images of handwritten digits from the MNIST dataset and learning the solution operators of partial differential equations, such as Burgers' equation and the radiative transfer equation. The potential for connections with artificial neural networks and other machine learning methods is also discussed.
- comments
- 12 pages, 8 figures
- fields
- Numerical Analysis