1: \begin{abstract}

2: 

3: We propose a Langevin diffusion-based algorithm 

4: for non-convex optimization and sampling 

5: on a product manifold of spheres. 

6: Under a logarithmic Sobolev inequality, 

7: we establish a guarantee for finite iteration 

8: convergence to the Gibbs distribution 

9: in terms of Kullback--Leibler divergence. 

10: We show that with an appropriate temperature choice, 

11: the suboptimality gap to the global minimum is guaranteed to be 

12: arbitrarily small with high probability.  

13: 

14: As an application, we analyze the proposed Langevin algorithm

15: for solving the Burer--Monteiro relaxation 

16: of a semidefinite program (SDP). 

17: In particular, we establish a logarithmic Sobolev inequality 

18: for the Burer--Monteiro problem when there are no spurious local minima; 

19: hence implying a fast escape from saddle points. 

20: Combining the results, we then provide a global optimality guarantee 

21: for the SDP and the Max-Cut problem. 

22: More precisely, we show the Langevin algorithm achieves 

23: $\epsilon$-multiplicative accuracy with high probability in 

24: $\widetilde{\Omega}( n^2 \epsilon^{-3} )$ iterations, 

25: where $n$ is the size of the cost matrix. 

26: 

27: \end{abstract}

28: