1: \begin{abstract}

2: 	Decentralized optimization with orthogonality constraints

3: 	is found widely in scientific computing and data science.

4: 	Since the orthogonality constraints are nonconvex,

5: 	it is quite challenging to design efficient algorithms.

6: 	Existing approaches leverage the geometric tools from Riemannian optimization

7: 	to solve this problem at the cost of high sample and communication complexities.

8: 	To relieve this difficulty, 

9: 	based on two novel techniques that can waive the orthogonality constraints,

10: 	we propose a variance-reduced stochastic gradient tracking (VRSGT) algorithm

11: 	with the convergence rate of $O(1 / k)$ to a stationary point.

12: 	To the best of our knowledge, 

13: 	VRSGT is the first algorithm for decentralized optimization with orthogonality constraints

14: 	that reduces both sampling and communication complexities simultaneously.

15: 	In the numerical experiments, 

16: 	VRSGT has a promising performance in a real-world autonomous driving application.

17: 	

18: 	

19: \end{abstract}

20: