摘要： In the context of future intelligent optical networks, dedicated learning techniques can be employed to monitor physical system parameters with a guaranteed accuracy. In this work, we investigate a method that establishes the link between input parameter uncertainties and the overall performance uncertainty. To this end, neglecting stochastic effects and focusing on the input parameters of a simplified Gaussian noise model version, we employ uncertainty propagation to evaluate the overall performance uncertainty from input parameter uncertainties, and we propose a simple way to link performance uncertainty to margins. With this method, as opposed to direct performance monitoring, it is possible to trace, in a predictable way, the path from the cause (input parameter uncertainties) to the effect (performance uncertainty) and to the additional network-level consequences (performance margins). We briefly review methods used in the literature to set margins in classical systems, and we show how all methods can be unified by means of the correlation between input parameters. By quantifying the impact of input parameter correlations, we further discuss the margins that can be saved if input parameters are partially correlated or uncorrelated, compared to a scenario in which parameters are fully correlated. We finally illustrate the separate impact of each parameter on performance uncertainty, and we briefly discuss their order of importance as a function of the system operating regime and propagated distance.