Sparse phase retrieval aims to recover a $k$-sparse signal from $m$ phaseless observations, raising the fundamental question of the minimum number of samples required for accurate recovery. As a classical non-convex optimization problem, sparse phase retrieval algorithms are typically composed of two stages: initialization and refinement. Existing studies reveal that the sampling complexity under Gaussian measurements is largely determined by the initialization stage. In this paper, we identify commonalities among widely used initialization algorithms and introduce key extensions to improve their performance. Building on this analysis, we propose a novel algorithm, termed Generalized Exponential Spectral Pursuit (GESP). Theoretically, our results not only align with existing conclusions but also demonstrate enhanced generalizability. In particular, our theoretical findings coincide with prior results under certain conditions while surpassing them in others by providing more comprehensive guarantees. Furthermore, extensive simulation experiments validate the practical effectiveness and robustness of GESP, showcasing its superiority in recovering sparse signals with reduced sampling requirements.
翻译:暂无翻译