Max vs Min: Tensor Decomposition and ICA with nearly Linear Sample Complexity