Assimilative Causal Inference

Causal inference is fundamental across scientific disciplines, yet existing methods struggle to capture instantaneous, time-evolving causal relationships in complex, high-dimensional systems. In this paper, assimilative causal inference (ACI) is developed, which is a methodological framework that leverages Bayesian data assimilation to trace causes backward from observed effects. ACI solves the inverse problem rather than quantifying forward influence. It uniquely identifies dynamic causal interactions without requiring observations of candidate causes, accommodates short datasets, and, in principle, can be implemented in high-dimensional settings by employing efficient data assimilation algorithms. Crucially, it provides online tracking of causal roles that may reverse intermittently and facilitates a mathematically rigorous criterion for the causal influence range, revealing how far effects propagate. The effectiveness of ACI is demonstrated by complex dynamical systems showcasing intermittency and extreme events. ACI opens valuable pathways for studying complex systems, where transient causal structures are critical.

💡 Research Summary

This paper introduces a novel framework called Assimilative Causal Inference (ACI) that leverages Bayesian data assimilation to infer causal relationships by tracing causes backward from observed effects. Traditional causal inference methods—such as Granger causality, transfer entropy, and model‑based information transfer—treat causality as a forward problem: they assess how a candidate cause influences future observations. In contrast, ACI formulates causality as an inverse problem. By running a stochastic dynamical model forward to generate a prior distribution of the system state and then assimilating observed data, two posterior distributions are obtained: a smoothing posterior that incorporates the entire observation window (past and future) and a filtering posterior that uses only past observations up to the current time. The key diagnostic is the relative entropy (Kullback‑Leibler divergence) between these two posteriors. If the smoothing posterior is more concentrated (i.e., its entropy is lower) than the filtering posterior, the future observations have reduced uncertainty about the current state of a candidate cause, indicating a causal influence at that time point.

The method is mathematically rigorous: the relative entropy accounts for both mean shifts and covariance changes, is invariant under nonlinear coordinate transformations, and can be computed efficiently with standard assimilation algorithms (Kalman filter, ensemble Kalman filter, particle filter, variational Bayes, etc.). This makes ACI scalable to high‑dimensional systems where classic information‑theoretic measures become intractable.

A second major contribution is the definition of the Causal Influence Range (CIR). CIR quantifies how far in space or time the effect of a cause propagates at any given instant. Two versions are proposed: a subjective CIR based on the decay of posterior mass below a chosen probability threshold, and an objective CIR derived from a mathematically justified entropy‑reduction criterion. Both are obtained directly from the assimilation output, eliminating the need for ad‑hoc thresholds.

The framework also handles additional non‑target variables through a conditional ACI formulation. By assigning infinite prior uncertainty to unobserved variables, their influence on the target cause’s posterior is effectively marginalized, allowing robust inference even when potential confounders are not measured.

The authors demonstrate ACI on two challenging nonlinear systems. The first is a dyadic model that exhibits intermittent extreme events; ACI successfully tracks rapid reversals of causal direction around bursts and provides a time‑varying CIR that expands during events and contracts during quiescent periods. The second application is a stochastic model of the El Niño‑Southern Oscillation (ENSO). Using both synthetic and real sea‑surface temperature data, ACI identifies the Pacific Ocean region that drives ENSO variability, captures regime switches, and yields CIR estimates that align with known teleconnection patterns. In both cases, ACI outperforms Granger causality and transfer entropy, which either miss the transient reversals or require long stationary datasets.

Overall, ACI offers several distinct advantages: (1) it requires observations only of the effect variable, making it suitable when candidate causes are unmeasured or partially observed; (2) it works with short time series because the inference relies on the information gain from future observations rather than long historical correlations; (3) it provides an online, time‑resolved causal map that can capture intermittent, regime‑dependent interactions; (4) it defines a principled, quantitative measure of causal reach (CIR); and (5) it integrates seamlessly with existing data‑assimilation infrastructure, enabling application to high‑dimensional geophysical, neuroscientific, or economic models. The paper concludes that assimilative causal inference opens a new pathway for studying complex dynamical systems where transient causal structures are essential, and it suggests future extensions such as incorporating observation noise, adaptive thresholding for CIR, and coupling with machine‑learning surrogates for even larger-scale problems.