For-all Sparse Recovery in Near-Optimal Time

Abstract: An approximate sparse recovery system in $\ell_1$ norm consists of parameters $k$, $\epsilon$, $N$, an $m$-by-$N$ measurement $\Phi$, and a recovery algorithm, $\mathcal{R}$. Given a vector, $\mathbf{x}$, the system approximates $x$ by $\widehat{\mathbf{x}} = \mathcal{R}(\Phi\mathbf{x})$, which must satisfy $|\widehat{\mathbf{x}}-\mathbf{x}|_1 \leq (1+\epsilon)|\mathbf{x}-\mathbf{x}_k|_1$. We consider the 'for all' model, in which a single matrix $\Phi$, possibly 'constructed' non-explicitly using the probabilistic method, is used for all signals $\mathbf{x}$. The best existing sublinear algorithm by Porat and Strauss (SODA'12) uses $O(\epsilon^{-3} k\log(N/k))$ measurements and runs in time $O(k^{1-\alpha}N^\alpha)$ for any constant $\alpha > 0$. In this paper, we improve the number of measurements to $O(\epsilon^{-2} k \log(N/k))$, matching the best existing upper bound (attained by super-linear algorithms), and the runtime to $O(k^{1+\beta}\textrm{poly}(\log N,1/\epsilon))$, with a modest restriction that $\epsilon \leq (\log k/\log N)^{\gamma}$, for any constants $\beta,\gamma > 0$. When $k\leq \log^c N$ for some $c>0$, the runtime is reduced to $O(k\textrm{poly}(N,1/\epsilon))$. With no restrictions on $\epsilon$, we have an approximation recovery system with $m = O(k/\epsilon \log(N/k)((\log N/\log k)^\gamma + 1/\epsilon))$ measurements.