Cause and effect lie at the heart of science: neuroscientists want to know which neurons drive a behavior, climatologists which factors dominate global temperature, biologists which genes regulate one another. Ideally, such questions are settled by experiment – intervene on a variable and watch how the system responds. But in many real-world complex systems, from the global climate to the human brain, deliberate intervention is prohibitively costly, technically infeasible, or simply impossible.
Now, reporting in The Innovation, a research team led by Professor Wei Lin of Fudan University, Professor Luonan Chen of the Chinese Academy of Sciences, and Professor Kazuyuki Aihara of the University of Tokyo has shown that genuine, intervention-level causality can be identified from observational data alone. Their new framework, interventional dynamical causality (IntDC), comes with a computational criterion – interventional embedding entropy (IEE) – that measures causal strength directly from recorded time series, without any knowledge of the system's dynamical equations and without performing a single experiment.
The new work distinguishes two levels of causality that existing methods tend to blur. Constructive dynamical causality –– the target of classical tools such as Granger causality, transfer entropy, and convergent cross mapping –– asks whether the dynamics of one variable depend on another. Such methods, however, are vulnerable to confounders: when a third factor drives both X and Y, they may wrongly report that X causes Y. Interventional dynamical causality adopts the stricter standard of modern causal inference: X counts as a cause of Y only if an external intervention on X would change Y.
Intervention is the gold standard of causal determination in experimental science, yet most important systems cannot be experimented on. The team's solution draws on a classic idea from dynamical systems theory: delay embedding, the theorem that the full state of a complex system can be reconstructed from the past behavior of just one of its measurements. The researchers' key insight is that if X exerts intervention-level influence on Y, then the observational record of Y already contains the imprint of X's perturbations. By reconstructing the system's state space from delay embeddings and defining a new information measure – the interventional embedding entropy – IEE quantifies how much intervened causal information flows between variables, allowing confounder-free causal structures to be recovered from passive observations.
Benchmarks confirmed the method's power. In classic logistic dynamical systems with a common driver Z, many traditional approaches produce spurious X→Y links; IEE correctly rejects them. The method also proved robust to data length, noise, embedding dimension, and parameter choice across logistic dynamics, coupled Hénon maps, and chaotic neural networks.
The real-world applications span the life sciences. Using calcium-imaging recordings from the nematode Caenorhabditis elegans, IEE reconstructed functional neural connectomes and quantified causal influence between brain regions. Applied to epidemic data, it identified transmission relationships between regions during Japan's COVID-19 outbreak. And in circadian biology, it rebuilt the regulatory network surrounding key clock genes, pinpointing the critical regulators.
The framework offers three practical advantages: it measures intervention-level causality in systems that cannot be manipulated; it quantifies causal strength in nonlinear, nonseparable systems in a comparable, rankable way; and it supports reconstruction of causal regulatory networks directly from data. As large-scale data and artificial intelligence reshape how science handles complex systems, the authors expect that combining dynamical causal theory with machine learning will further sharpen data-driven causal discovery – turning observation alone into a window on cause and effect.
