When is Distributed as Good as Centralized?
Jose' M. F. Moura
From networks of social agents, interacting locally but aspiring to global understanding, to physical networked systems like the power grid, where distributed SCADAs could become candidates to replace centralized supervisory control and data acquisition systems, we can ask if and when a distributed algorithm offers performance guarantees similar to those of a centralized solution. We consider consensus+innovations algorithms that combine local cooperation among agents (in-network processing) with local exchanges with the external world (sensing). To understand these mixed scale algorithms we establish their path behavior and the exponential decay rates of the probability of rare events (large deviations principle.) These algorithms need careful design, in particular, how to weigh the consensus and innovations terms. We illustrate relevant tradeoffs among network parameters (e.g., rate of diffusion, communication and sensing signal-to-noise ratios,) and determine that, under broad conditions, yes, distributed can be as good as centralized.