Line spectral estimation or frequency estimation is a fundamental problem in statistical signal processing. It aims at representing a natural signal as superposition of several sinusoidal waves. The use of multichannel data for line spectral estimation arises in applications such as array processing, radar, structural health monitoring, wireless communications, and more. In the past decade, sparse representation and compressed sensing techniques have demonstrated their superiority in flexibility, accuracy and robustness in comparison with conventional nonparametric and parametric methods. However, their performance is usually questionable due to approximation introduced and lack of theoretical analysis. In this talk, we introduce the most recent developments in this area using atomic norm. As continuous analogs of L1 norm minimization approaches, atomic norm methods exploit signal sparsity, work directly in the continuous parameter domain, can be implemented using convex optimization, and have provable theoretical guarantees. The role of multiple channels is discussed by analysing their worst-case and average-case performances.