Observability Collapse: A Mechanism for the Failure of Early Warning Signals

This record contains the preprint: Observability Collapse: A Mechanism for the Failure of Early Warning Signals Early warning signals based on critical slowing down (e.g., rising variance and autocorrelation) are widely used to anticipate transitions in complex systems. However, empirical studies often report weak, absent, or reversed patterns prior to transition. This work proposes a dynamical mechanism—observability collapse—in which the measurable signal decreases as the system approaches an attractor when the observation function is state-dependent. Under these conditions, early warning signals may be suppressed despite increasing latent instability. The mechanism is formalized in a stochastic dynamical framework and demonstrated through simulation, including an ablation showing that signal suppression arises from the observation function rather than from system dynamics. An exploratory empirical analysis using experience sampling data does not recover the predicted signature, consistent with limitations in measurement and timescale. Code to reproduce all simulations and analyses is available at: https://github.com/algbz/observability-collapse