Decentralized asynchronous optimization with DADAO allows decoupling and acceleration (Q7308313)

This work proposes an algorithm that seems to be the first decentralized, accelerated, asynchronous, first-order method for minimizing a sum of \(L\)-smooth and \(\mu\)-strongly convex functions distributed over a network of size \(n\), that unifies several recent approaches. Technical results on \(\epsilon\)-precision reveal very competitive theoretical speed guarantees of the proposed method, that outperforms the state-of-the-art algorithms in the field. Moreover, an SDP relaxation is proposed for deriving the gossip rate of each edge, minimizing the total number of communications for a given graph. Some numerical results and direct comparisons illustrate the theoretical gains of the paper.