1: \begin{abstract}

2: Privacy-preserving distributed distribution comparison measures the distance between the distributions whose data are scattered across different agents in a distributed system and cannot be shared among the agents. 

3: In this study, we propose a novel decentralized entropic optimal transport (EOT) method, which provides a \textit{privacy-preserving} and \textit{communication-efficient} solution to this problem with theoretical guarantees. 

4: In particular, we design a mini-batch randomized block-coordinate descent (MRBCD) scheme to optimize the decentralized EOT distance in its dual form.

5: The dual variables are scattered across different agents and updated locally and iteratively with limited communications among partial agents.

6: The kernel matrix involved in the gradients of the dual variables is estimated by a distributed kernel approximation method, and each agent only needs to approximate and store a sub-kernel matrix by one-shot communication and without sharing raw data. 

7: We analyze our method's communication complexity and provide a theoretical bound for the approximation error caused by the convergence error, the approximated kernel, and the mismatch between the storage and communication protocols. 

8: Experiments on synthetic data and real-world distributed domain adaptation tasks demonstrate the effectiveness of our method.

9: \end{abstract}

10: