A fundamental problem in forecasting the future of ecological systems is the ominous possibility of non-linear responses to perturbations. Multiple stable states are one possible outcome of such things, but even if that doesn't happen, non-linearity makes a mockery of predictions from our standard theoretical tools.
Even picking up thresholds in data is hard - I've used maximum likelihood estimation of split line models to do this, but you have to decide in advance how many different "thresholds" to put into the models. Derek Sonderegger and colleagues present a different way of identifying thresholds in this weeks Frontier's in Ecology and the Environment - using a method called SiZer that examines changes in the derivatives of smooth curves fit to the data with different bandwidths. This is way cool - makes nice pictures too!
Sonderegger et al. (2009) Using SiZer to detect thresholds in ecological data. 7(4):190-195