学术文献库

A broad range of natural and social systems from human microbiome to financial markets can go through critical transitions, where the system suddenly collapses to another stable configuration. Critical transitions can be unexpected, with potentially catastrophic consequences. Anticipating them early and accurately can facilitate controlled system manipulation and mitigation of undesired outcomes. Obtaining reliable predictions have been difficult, however, as often only a small fraction of the relevant variables can be monitored, and even minor perturbations can induce drastic changes in fragile states of a complex system. Data-driven indicators have been proposed as an alternative to prediction and signal an increasing risk of forthcoming transitions. Autocorrelation and variance are examples of generic indicators that tend to increase at the vicinity of an approaching tipping point across a range of systems. An important shortcoming in these and other widely studied indicators is that they deal with simplified one-dimensional representations of complex systems. Here, we demonstrate that a probabilistic data aggregation strategy can provide new ways to improve early warning detection by more efficiently utilizing the available information from multivariate time series. In particular, we consider a probabilistic variant of a vector autoregression model as a novel early warning indicator and argue that it has theoretical advantages related to model regularization, treatment of uncertainties, and parameter interpretation. We evaluate the performance against alternatives in a simulation benchmark and show improved sensitivity in EWS detection in a common ecological model encompassing multiple interacting species.