In this talk we will focus on the high dimensional linear regression problem. The goal is to recover a hidden k-sparse binary vector \beta under n noisy linear observations Y=X\beta+W where X is an n \times p matrix with iid N(0,1) entries and W is an n-dimensional vector with iid N(0,\sigma^2) entries. In the literature of the problem, an apparent asymptotic gap is observed between the optimal sample size for information-theoretic recovery, call it n*, and for computationally efficient recovery, call it n_alg.

We will discuss several new contributions on studying this gap. We first identify tightly the information limit of the problem using a novel analysis of the Maximum Likelihood Estimator (MLE) performance. Furthermore, we establish that the algorithmic barrier n_alg coincides with the phase transition point for the appearance of a certain Overlap Gap Property (OGP) over the space of k-sparse binary vectors. The presence of such an Overlap Gap Property phase transition, which originates in spin glass theory, is known to provide evidence of an algorithmic hardness. Finally, we show that in the extreme case where the noise level is zero, i.e. \sigma=0, the computational-statistical gap closes by proposing an optimal polynomial-time algorithm using the Lenstra-Lenstra-Lov\'asz lattice basis reduction algorithm.

This is joint work with David Gamarnik.