Distributed structured computation

We devise a novel source encoding technique for distributed matrix multiplication. Our

approach relies on devising nonlinear mappings of distributed sources, followed by a structured linear encoding scheme introduced by Körner and Marton. For different computation scenarios, we contrast our findings on the achievable sum rate with the state of the art to demonstrate the possible savings in compression rate. When the sources have specific correlation structures, it is possible to achieve unbounded gains, as indicated by the analysis and numerical simulations. In the framework of coded matrix multiplication, we explore the tradeoff between the worker storage constraint, the end-to-end communication and computation costs, and the recovery threshold. We showcase the

gains versus existing codes and how we can expand the cost tradeoff space via interpolating between various codes.