1: \begin{abstract}

2: In this work, we study the tensor ring decomposition and its

3: associated numerical algorithms. We establish a sharp transition

4: of algorithmic difficulty of the optimization problem as the bond

5: dimension increases: On one hand, we show the existence of spurious

6: local minima for the optimization landscape even when the tensor ring

7: format is much over-parameterized, \textit{i.e.}, with bond dimension

8: much larger than that of the true target tensor.  On the other hand,

9: when the bond dimension is further increased, we establish one-loop

10: convergence for alternating least square algorithm for the tensor ring

11: decomposition. The theoretical results are complemented by numerical

12: experiments for both local minima and the one-loop convergence for

13: the alternating least square algorithm.

14: \end{abstract}

15: