Forecasts of chaotic systems like the atmosphere become contaminated, then dominated by noise unrelated to the true state of the system. Ensemble forecasting is designed to sample the space of forecast error. At most centers, integrations from perturbed initial conditions have augmented or replaced higher resolution control forecasts started from the best initial condition. Beyond alternative scenarios, ensembles provide a wide range of probabilistic and other products.Random perturbations have a statistically equal projection in each independent phase space direction. Hence in the high dimensional space of atmospheric dynamics, even if statistically indistinguishable from error fields, perturbations have a very small projection on the actual realization of error; the bulk of the variance adds noise in other directions. This results in a cloud of solutions not around, but further displaced from reality. Initial error is doubled, causing a 20-hour drop in forecast skill, equivalent to using NWP output from 8 years ago. This behavior is observed in operational, perfect, and statistically simulated ensembles, suggesting it is not caused by methodological problems. Instead, the failure is due to fundamental limitations in sampling the multidimensional space of atmospheric dynamics. 

集合预报