Metric-based indicators
Results from BDS tests and their partial autocorrelation functions ACF show that we can reject the null hypothesis that the remaining time series residuals after detrending are independent and identically distributed, which is typical of a system approaching a critical transition (Appendix S1, Table S3 and Fig. S1). ACF also showed that during the activity period, body temperature followed a circadian cyclicity, which was not apparent during hibernation. Conditional heteroskedasticity (CH) was erratic and did not clearly anticipate bifurcation between states (Appendix S1, Fig. S2). DDJ metrics were noisy when plotted against time, although they were suitable indicators of resilience for flickering data: resilience decreased as conditional variance, diffusion and jump intensity increased (Appendix S1, Fig. S3). Generic early warning signals (EWS) were also noisier for the transition to hibernation, likely because the time series started just before dormice began to show an increasing frequency of torpor bouts (Fig. 2). In general, there was an increase in the autocorrelation at lag-1 and in variance before the two transitions, whereas skewness decreased, likely due to the increase in excursions of body temperature toTe over this period. Generic EWS showed similar performance for the time series encompassing the transitions and the whole time series (Appendix S1, Fig. S4 and S5, respectively). Sensitivities of all generic EWS tested in our study were low, i.e. results were robust regardless of different choices on bandwidth and size of the rolling window (Appendix S1, Fig. S6).