1: \begin{abstract}

2: Given a graphical model (GM),

3: computing its partition function is the most essential inference task,

4: but it is computationally intractable in general.

5: To address the issue, iterative approximation algorithms exploring certain local structure/consistency of GM

6: have been investigated as popular choices in practice.

7: However, due to their local/iterative nature, they often output poor approximations or even do not converge, e.g.,

8: in low-temperature regimes (hard instances of large parameters).

9: To overcome the limitation, we propose a novel approach utilizing the global spectral feature of GM.

10: Our contribution is two-fold:

11: (a) we first propose a fully polynomial-time approximation scheme (FPTAS) for approximating the partition function of GM 

12: associating with a low-rank coupling matrix;

13: (b) for general high-rank GMs,

14: we design a spectral mean-field scheme utilizing (a) as a subroutine, where

15: it approximates a high-rank GM into a product of rank-1 GMs for an efficient approximation of the partition function.

16: The proposed algorithm is more robust in its running time and accuracy than prior methods, i.e., 

17: neither suffers from the convergence issue nor depends on hard local structures, as demonstrated in

18: our experiments.

19: \end{abstract}

20: