L2/L2-foreach sparse recovery with low risk

In this paper, we consider the "foreach" sparse recovery problem with failure probability $p$. The goal of which is to design a distribution over $m \times N$ matrices $\Phi$ and a decoding algorithm $\algo$ such that for every $\vx\in\R^N$, we have the following error guarantee with probability at least $1-p$ $$\|\vx-\algo(\Phi\vx)\|_2\le C\|\vx-\vx_k\|_2,$$ where $C$ is a constant (ideally arbitrarily close to 1) and $\vx_k$ is the best $k$-sparse approximation of $\vx$. Much of the sparse recovery or compressive sensing literature has focused on the case of either $p = 0$ or $p = \Omega(1)$. We initiate the study of this problem for the entire range of failure probability. Our two main results are as follows: \begin{enumerate} \item We prove a lower bound on $m$, the number measurements, of $\Omega(k\log(n/k)+\log(1/p))$ for $2^{-\Theta(N)}\le p <1$. Cohen, Dahmen, and DeVore \cite{CDD2007:NearOptimall2l2} prove that this bound is tight. \item We prove nearly matching upper bounds for \textit{sub-linear} time decoding. Previous such results addressed only $p = \Omega(1)$. \end{enumerate} Our results and techniques lead to the following corollaries: (i) the first ever sub-linear time decoding $\lolo$ "forall" sparse recovery system that requires a $\log^{\gamma}{N}$ extra factor (for some $\gamma<1$) over the optimal $O(k\log(N/k))$ number of measurements, and (ii) extensions of Gilbert et al. \cite{GHRSW12:SimpleSignals} results for information-theoretically bounded adversaries.