Compressed Communication Framework for Decentralized Nonconvex Optimization
July 3, 2026
This framework unifies decentralized stochastic subgradient methods that use unbiased or contractive compression with error compensation. It establishes global convergence for nonsmooth, nonconvex objective functions that lack Clarke regularity by relating iterates to continuous-time differential inclusions.
HOW THIS AFFECTS YOU
●
researcherYou can leverage these convergence guarantees when designing decentralized training protocols with limited communication bandwidth.