Cyber-physical technologies enable event-driven applications, which monitor in real-time the occurrence of certain inherently stochastic incidents. Those technologies are being widely deployed in cities around the world and one of their critical aspects is energy consumption, as they are mostly battery powered. The most representative examples of such applications today is smart parking. Since parking sensors are devoted to detect parking events in almost-real time, strategies like data aggregation are not well suited to optimize energy consumption. Furthermore, data compression is pointless, as events are essentially binary entities. Therefore, this paper introduces the concept of Lean Sensing, which enables the relaxation of sensing accuracy at the benefit of improved operational costs. To this end, this paper departs from the concept of instantaneous randomness and it explores the correlation structure that emerges from it in complex systems. Then, it examines the use of this system-wide aggregated contextual information to optimize power consumption, thus going in the opposite way; from the system-level representation to individual device power consumption. The discussed techniques include customizing the data acquisition to temporal correlations (i.e, to adapt sensor behavior to the expected activity) and inferring the system-state from incomplete information based on spatial correlations. These techniques are applied to real-world smart-parking application deployments, aiming to evaluate the impact that a number of system-level optimization strategies have on devices power consumption.