1: \begin{abstract}

2:   In this paper we propose an efficient method to compress a high

3:   dimensional function into a tensor ring format, based on alternating

4:   least-squares (ALS). Since the function has size exponential in $d$

5:   where $d$ is the number of dimensions, we propose efficient sampling

6:   scheme to obtain $O(d)$ important samples in order to learn the

7:   tensor ring. Furthermore, we devise an initialization method for ALS

8:   that allows fast convergence in practice. Numerical examples show

9:   that to approximate a function with similar accuracy, the tensor

10:   ring format provided by the proposed method has less parameters than

11:   tensor-train format and also better respects the structure of the

12:   original function.

13: \end{abstract}

14: