arXiv · 2611.22531
Inapproximability of L1->L2 Subspace Distortion
Abstract
We show, assuming NP is not contained in randomized quasipolynomial time, almost polynomial factor inapproximability for maximizing the $\ell_2/\ell_1$ ratio over a real subspace -- a problem also known as subspace distortion. This is a central problem in compressed sensing but also has connections to many areas such as real error correcting codes, dictionary learning, hypercontractive operator norms, small-set expansion, quantum information and random matrix theory. Distortion seems to be at the heart of building inapproximability theory for continuous optimization. Extending our results to the case of $\ell_2/\ell_p$ for all $p<2$ would lead to a near characterization of the convex sets over which one can efficiently and approximately maximize an arbitrary quadratic form. A recurring obstruction appears in hardness considerations of distortion -- how does one use a sparsity measure to test proximity to an exponential sized discrete set over the reals? This issue also appears in a host of other problems that involve measures of sparsity, shortness, density or expansion. We believe hardness of distortion is an intermediate step to understanding the complexity of small-set expansion and quantum separability, both of which are important directions of research. The main insight in our hardness reduction is the construction of a randomized affine linear map $T:\mathbb{R}^n\to \mathbb{R}^m$ such that the distortion of $Tx$ measures proximity to the hypercube $\{\pm 1\}^{n}$. Crucial to our construction is subsampling $x$ at *all* dyadic sample sizes, which gives an inverse polylogarithmic gap. This is a highly non-local tester, and thus departs from the local testing paradigm in PCP theory. Our reduction is also quite succinct, and so it suggests a rather simple distribution of subspaces for which approximately finding a planted sparse vector is intractable.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Vijay Bhattiprolu. 2026-11-17. Inapproximability of L1->L2 Subspace Distortion. https://arxiv.science/abs/2611.22531
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.