A "Propagative" Approach to Sensitivity Analysis
Large-scale simulations often involve huge numbers of parameters, making it prohibitive to run more than a tiny fraction of all potentially relevant cases. Sensitivity analysis attempts to show how responsive the results of a simulation are to changes in its parameters: this is an important tool for promoting confidence in a simulation and making its results credible. However, the computational cost of traditional approaches to sensitivity analysis prevents its use in many cases. The authors show that this cost is logically unnecessary and can be largely avoided by propagating and combining sensitivities during a computation, rather than recomputing them. They describe this "propagative" approach to sensitivity analysis and present the algorithm they have implemented to explore its potential.