Fitting Gaussian vectors in normed spaces

Abstract

Given a family of $N$ i.i.d. Gaussian points in a $d$-dimensional normed space, what is the probability that some linear transformation maps it to the unit sphere? The case of a Euclidean space is known as the ellipsoid fitting problem and has recently attracted major attention. We initiate a study of this fitting problem in a general normed space, both asymptotically and in a fixed dimension. We specifically focus on the space $\ell_\infty^d$, where a combinatorial reformulation allows us to show that $cd^2/\log d$ points can be fitted with high probability. We illustrate that the geometry of the normed space is reflected in the fitting problem by showing that in dimension $2$, the spaces $\ell_2^2$ and $\ell_\infty^2$ are characterized by having extremal fittability properties.