Statistical guarantees for sparse deep learning