RESTRICTED ISOMETRY CONSTANTS WHERE l p SPARSE RECOVERY CAN FAIL FOR 0 < p ≤ 1

We investigate conditions under which the solution of an underdetermined linear system with minimal lp norm, 0 < p ≤ 1, is guaranteed to be also the sparsest one. Our results highlight the pessimistic nature of sparse recovery analysis when recovery is predicted based on the restricted isometry constants (RIC) of the associated matrix. We construct matrices with RIC δ2m arbitrarily close to 1/ √ 2 ≈ 0.717 where sparse recovery with p = 1 fails for at least one m-sparse vector. This indicates that there is limited room for improving over the best known positive results of Foucart and Lai, which guarantee that l1-minimisation recovers all m-sparse vectors for any matrix with δ2m < 2(3 − √ 2)/7 ≈ 0.4531. Another consequence of our construction is that recovery conditions expressed uniformly for all matrices in terms of RIC must require that all 2m-column submatrices are extremely well conditioned (condition numbers less than 2.5). In contrast, we also construct matrices with δ2m arbitrarily close to one and δ2m+1 = 1 where l1-minimisation succeeds for any m-sparse vector. This illustrates the limits of RIC as a tool to predict the behaviour of l1 minimisation. These constructions are a by-product of tight conditions for lp recovery (0 ≤ p ≤ 1) with matrices of unit spectral norm, which are expressed in terms of the minimal singular values of 2m-column submatrices. The results show that, compared to l1-minimisation, lp-minimisation recovery failure is only slightly delayed in terms of the RIC values. Furthermore in this case the minimisation is nonconvex and it is important to consider the specific minimisation algorithm being used. We show that when lp optimisation is attempted using an iterative reweighted l1 scheme, failure can still occur for δ2m arbitrarily close to 1/ √ 2. Key-words: underdetermined linear system, sparse representation, overcomplete dictionary, compressed sensing, inverse problem, restricted isometry property, convex optimisation, nonconvex optimisation, iterative reweighted optimisation.