Classical sampling records the signal level at pre-determined time instances, usually uniformly spaced. An alternative implicit sampling model is to record the timing of pre-determined level crossings. Thus the signal dictates the sampling times but not the sampling levels. Logan's theorem provides sufficient conditions for a signal to be recoverable, within a scaling factor, from only the timing of its zero crossings. Unfortunately, recovery from noisy observations of the timings is not robust and usually fails to reproduce the original signal. To make the reconstruction robust this paper introduces the additional assumption that the signal is sparse in some basis. We reformulate the reconstruction problem as a minimization of a sparsity inducing cost function on the unit sphere and provide an algorithm to compute the solution. While the problem is not convex, simulation studies indicate that the algorithm converges in typical cases and produces the correct solution with very high probability.