[LG] On the Theoretical Understanding of Identifiable Sparse Autoencoders and Beyond[Peking University & MIT]arxiv.org