Abstract
Identifying causal relationships in nonlinear dynamical systems remains a significant challenge, as traditional statistical methods often fail to distinguish between direct and indirect links in the presence of feedback loops and synergistic variables. This paper presents MXMap, a two-phase framework that extends Partial Cross Mapping to a multivariate setting (multiPCM). By leveraging multivariate state-space reconstruction, MXMap effectively prunes indirect causal links and identifies complex interactions, including cycles, in high-dimensional systems. Validation on simulated Lotka-Volterra models and real-world ERA5 meteorological data demonstrates that MXMap outperforms existing baselines in both accuracy and consistency.
Introduction
Understanding the causal architecture of complex systems—such as global climate patterns or multi-species ecosystems—is a fundamental goal of scientific inquiry. However, traditional causal inference tools like Granger Causality often rely on the assumption that a cause contains unique information to predict the future of its effect. In nonlinear, coupled dynamical systems, this assumption frequently breaks down as variables become chaotic or synergistic.
While Convergent Cross Mapping (CCM) introduced a breakthrough by using state-space reconstruction to detect causality, it is prone to "transitivity" errors. For example, in a chain $A \to B \to C$, CCM may incorrectly identify a direct link from $A$ to $C$. While Partial Cross Mapping (PCM) was developed to address this via conditioning, it was previously limited to univariate data. MXMap addresses these limitations by introducing a multivariate framework capable of refining dense causal graphs in high-dimensional settings.
Methodology
MXMap operates through a two-phase process designed to move from a broad set of potential interactions to a refined, accurate causal graph.
Phase 1: Initial Graph Construction
The framework begins by performing exhaustive pairwise CCM tests between all variables in the system. This phase captures all potential interactions, resulting in an initial "dense" causal graph. This graph typically includes both direct causal links and "fake" indirect links caused by transitivity.
Phase 2: Pruning via multiPCM
The core innovation of MXMap is the multiPCM algorithm. For every suspected link in the initial graph, the algorithm identifies all possible intermediate paths. It then utilizes multivariate state-space reconstruction (multiSSR)—stacking multiple time-lagged variables into a single state-space—to determine if the causal information between a parent and child remains significant after conditioning on the rest of the system. If the direct information transfer is negligible when accounting for other variables, the link is pruned as indirect.
Experimental Evaluation
Simulated Systems
The framework was validated using simulated species interaction systems based on Lotka-Volterra competition models. These simulations ranged from 3 to 7 variables and included various topologies such as chains, "immoralities" (two causes, one effect), and cycles. MXMap demonstrated exceptional precision, achieving a Structural Hamming Distance (SHD) of 0 (perfect reconstruction) in several scenarios where baselines like Granger Causality and DYNOTEARS failed significantly, sometimes reaching SHDs as high as 10.
Real-World Meteorological Data
To test real-world applicability, the authors applied MXMap to ERA5 reanalysis weather data for the Montreal region (1981–2023). Focusing on winter months, the study targeted the established causal chain: Total Cloud Water ($tcw$) $\to$ Radiation ($rad$) $\to$ Near-Ground Temperature ($T_{2m}$). MXMap correctly identified these benchmark relationships with an F1 score of 0.80, significantly outperforming established methods such as PCMCI (0.67) and VAR-LiNGAM (0.0).
Key Insights
- Handling Cycles: Unlike many causal discovery methods that assume a Directed Acyclic Graph (DAG), MXMap is explicitly designed to handle feedback loops, which are ubiquitous in natural systems.
- Resilience to Mirage Correlations: Nonlinear systems often produce "mirage correlations" that appear and disappear over time. MXMap shows greater consistency across different time segments compared to regression-based methods like RESIT.
- Multivariate Conditioning: By using a multivariate set for conditioning, the model accounts for multiple paths of indirect influence simultaneously, which is essential for high-dimensional data.
Discussion and Limitations
While MXMap represents a significant advancement, it possesses certain limitations. Its runtime complexity is $O(n^2)$ due to pairwise processing; while efficient for the 8-variable systems tested (approx. 40 seconds), it may require optimization for systems with hundreds of variables. Additionally, the method assumes the system settles into a stable "attractor," meaning it may struggle with purely stochastic or transient systems. Future work will focus on adaptive parameter selection to automatically tune the lag ($\tau$) and embedding dimension ($E$), which currently require empirical selection.
Summary updated