Convergent Cross Mapping

Convergent cross mapping (CCM) is a method for detecting causality in complex dynamical systems, developed by George Sugihara and colleagues in 2012. Unlike the directed acyclic graph framework of causal inference, which assumes that causes can be identified by intervening on variables and observing downstream effects, CCM detects causality by exploiting a property of nonlinear dynamical systems: variables that are causally coupled will share a common attractor in their reconstructed state spaces.

The core insight is that if X causes Y, then the historical record of Y contains information about the state of X. In a deterministic dynamical system, the state of Y is a function of the states of all its causal drivers, including X. Therefore, a sufficiently rich history of Y can be used to reconstruct — or "cross-map" — the state of X. If X does not cause Y, no such reconstruction is possible, because the state of Y contains no information about X.

The Method

CCM proceeds by reconstructing the state space of each variable using time-delay embedding (the method of Takens' theorem). For a time series of Y, one constructs a set of points in a high-dimensional space where each point is a vector of lagged values: Y(t), Y(t-τ), Y(t-2τ), ... . Takens' theorem guarantees that under mild conditions, this reconstructed space has the same topological properties as the original dynamical system — the same attractor, the same dimensionality, the same Lyapunov exponents.

To test whether X causes Y, CCM looks for convergent cross-mapping: as the length of the time series increases, the accuracy of predicting X from Y's reconstructed state space should improve. The convergence is key: in purely correlated systems (e.g., two variables driven by a common third variable), the cross-mapping accuracy does not improve with more data, because the shared information is static. In causally coupled systems, more data means better sampling of the attractor, which means better cross-mapping.

Why CCM Matters for Systems

CCM matters because it solves a problem that DAG-based causal inference cannot solve: detecting causality in systems where experiments are impossible and where feedback loops make the notion of "intervention" incoherent. In ecology, one cannot intervene on the Pacific climate system to test whether the El Niño-Southern Oscillation causes salmon population cycles. In economics, one cannot intervene on the stock market to test whether investor sentiment causes volatility. In neuroscience, one cannot intervene on a single neuron embedded in a recurrent network to test whether it causes a specific pattern of population activity.

In all these cases, CCM provides a way to detect causality from observational time-series data alone, without assuming separability, without breaking feedback loops, and without requiring the strong ignorability assumptions that DAG-based methods demand. The trade-off is that CCM requires nonlinear dynamical systems with sufficient complexity to produce a reconstructible attractor. It does not work for purely stochastic systems, for systems with too little data, or for systems where the causal coupling is too weak to leave a detectable signature on the attractor.

The Synthesis: CCM and Causal Inference

The synthesizer's reading is that CCM and Pearl's do-calculus are not competitors but complements, operating at different scales and with different assumptions. Pearl's framework is the right tool when the system can be decomposed into independent variables, when interventions are possible, and when the goal is to estimate the magnitude of a specific causal effect. CCM is the right tool when the system is a tightly coupled dynamical whole, when interventions are impossible, and when the goal is to detect whether a causal link exists at all.

The deeper synthesis is that both methods are trying to solve the same problem — how to read causality from data — but they make different commitments about what causality looks like. Pearl assumes causality is a structural relation between variables. CCM assumes causality is a dynamical relation between trajectories. The variables framework is appropriate when the system is modular. The trajectories framework is appropriate when the system is integrated. The mistake is to apply either where the other belongs.

Convergent cross mapping is the method for detecting causality in systems that do not let you pull them apart. It is the causal inference of the whole, not the part.