Anisotropic Gaussian approximation in $L_2(\mathbb{R}^2)$

TL;DR

This paper demonstrates that Gaussian mixtures can effectively approximate anisotropic structures in 2D, achieving error bounds comparable to optimal curvelet approximations by developing new analysis tools for $L_2$ error measurement.

Contribution

It provides the first rigorous theoretical results showing Gaussian mixtures' effectiveness in approximating anisotropic features in 2D, with algorithms and error bounds matching curvelet performance.

Findings

Gaussian mixtures can resolve anisotropic structures in 2D.

The approximation error in $L_1$-norm matches curvelet bounds.

New machinery is developed for $L_2$-norm error analysis.

Abstract

Let $\mathcal{D}$ be the dictionary of Gaussian mixtures: the functions created by affine change of variables of a single Gaussian in $n$ dimensions. $\mathcal{D}$ is used pervasively in scientific applications to a degree that practitioners often employ it as their default choice for representing their scientific object. Its use in applications hinges on the perception that this dictionary is large enough, and its members are local enough in space and frequency, to provide efficient approximation to "almost all objects of interest". However, and perhaps surprisingly, only a handful of concrete theoretical results are actually known on the ability to use Gaussian mixtures in lieu of mainstream representation systems. The present paper shows that, in 2D, Gaussian mixtures are effective in resolving anisotropic structures, too. In this setup, the "smoothness class" is comprised of 2D functions that are sparsely represented using curvelets. An algorithm for $N$-term approximation from (a small subset of) $\mathcal{D}$ is presented, and the error bounds are then shown to be on par with the errors of $N$-term curvelet approximation. The latter are optimal, essentially by definition. Our approach is based on providing effective approximation from $\mathcal{D}$ to the members of the curvelet system, mimicking the approach in arXiv:0802.2517 and arXiv:0911.2803 where the mother wavelets are approximated. When the error is measured in the $1$-norm, this adaptation of the prior approach, combined with standard tools, yields the desired results. However, handling the $2$-norm case is much more subtle and requires substantial new machinery: in this case, the error analysis cannot be solely done on the space domain: some of it has to be carried out on frequency. Since, on frequency, all members of $\mathcal{D}$ are centered at the origin, a delicate analysis for controlling the error there is needed.