Causal Fingerprints Reveal Hidden Mechanisms in Partially-Observed Dynamical Systems

Résumé

In partially-observed dynamical systems, distinct hidden mechanisms can produce similar observable summary statistics while explaining individual events differently. We introduce causal fingerprints: realization-dependent patterns that track how information about a hidden state is revealed through subsequent evolution. Within a prescribed dynamical model, we compare the filtering posterior, based on observations up to a given time, with the smoothing posterior, obtained after adding an expanding window of future observations. We quantify this retrospective information gain using Kullback-Leibl…

• #math.DS