Search papers, labs, and topics across Lattice.
This paper confirms the conjecture by Axiotis and Sviridenko regarding the limitations of polynomial-time algorithms in sparse convex optimization, specifically for least-squares objectives. By establishing a lower bound that relies on the randomized exact-volume Small-Set Expansion Hypothesis, the authors demonstrate that no randomized polynomial-time algorithm can achieve a certain sparsity condition while minimizing the least-squares error. The findings underscore the inherent challenges posed by the restricted condition number in sparse optimization, even in rational instances with full column rank matrices.
No randomized polynomial-time algorithm can overcome the condition-number barrier in sparse least-squares optimization, confirming a long-standing conjecture.
In [AS21], Axiotis and Sviridenko conjectured that the linear dependence on the restricted condition number in sparse convex optimization cannot be improved by a polynomial-time algorithm. We establish their conjectured lower bound for least-squares objectives, conditional on the randomized exact-volume Small-Set Expansion Hypothesis in the weighted regular-graph formulation of Raghavendra, Steurer, and Tulsiani [RST12]. Concretely, for every fixed $γ\in(0,1]$, there is no randomized polynomial-time algorithm that, with probability at least $2/3$, returns a vector $x$ such that, writing $s=\lVert x\rVert_0$, \[ \lVert Ax-b\rVert_2^2 \leq \min_{\lVert z\rVert_0\leq k}\lVert Az-b\rVert_2^2+\varepsilon \quad\text{and}\quad s=O\!\left(k\,κ_{s+k}^{\,1-γ}\right), \] where $κ_r$ is the restricted condition number at sparsity level $r$. The result holds even on rational instances with $A$ of full column rank. The proof was first obtained using a fully automated Gemini-based agentic system developed internally at Google. The authors have verified the proof and edited it for clarity of presentation.
CMU ML
Google Research