Inferring critical transitions in paleoecological time series with irregular sampling and variable time-averaging

Abstract

Abstract Many ecosystems have abruptly changed in the past and may again in the future, yet prediction and inference of mechanisms causing abrupt changes remains challenging. Critical transitions are one such mechanism, occurring when systems with alternative states cross a threshold. Such transitions are associated with a loss of resilience, often signaled by increasing variability or autocorrelation over time. However, critical transitions are difficult to distinguish from other causal mechanisms, and detection of resilience loss in sedimentary archives can be confounded by time-averaging and discontinuous sampling. Here, we simulate woodland-grassland regime shifts resulting from critical transitions and other mechanisms. We then test the diagnostic ability of two widely-used resilience indicators, standard deviation and autocorrelation time, after alterations common in sedimentary records: time averaging, discontinuous sampling, and varying sedimentation rates. Standard deviation—but not autocorrelation time—still distinguishes gradually forced critical transitions from other regime shifts when sedimentation rates are constant, and can be robust to abrupt changes in sedimentation rate. Unfortunately, shifts in standard deviation alone are rarely definitive evidence of critical transitions. Under exponential sedimentation regimes, which are common in younger upper-column sediments, neither resilience indicator is effective. Discontinuous sampling weakened the strength of resilience indicators. A demonstrative analysis of abrupt early Holocene deforestation recorded at Steel Lake, Minnesota showed signals consistent with resilience loss during early Holocene aridification. Hence, signals of resilience loss can be recovered from sedimentary archives, but efficacy varies among indicators and sedimentation regime. High-resolution and multi-proxy records remain essential to inferring causes, while process-based time series modeling such as this can be calibrated to systems of interest to explicitly test hypotheses about abrupt change causes.