DCatalyst: A Unified Accelerated Framework forDecentralized Optimization:

We designed and proposed a unified framework called DCatalyst to accelerate the decentralized algorithms for various objective functions. The complete analysis and comprehensive simulation results are provided.

Enhancing Convergence of Decentralized Gradient Tracking under the KL Property:

Leveraging the kurdyka-Lojasiewicz property, we enhanced the convergence analysis of gradient tracking algorithm. Notably, linear convergence is provably obtained for KL objective function with the exponent 1/2.

Parameter-free algorithms for decentralized optimization

We investigated the optimization algorithms on the decentralized settings that require no prior parametric knowledge on either objective functions or networks. Our main focus now is on the composite problems.