Causal Discovery via State-Space Reconstruction

24 concepts21 papers1 proposal

In complex, non-linear systems—ranging from the global climate to the firing of neurons—traditional statistical tools often fail to distinguish between coincidence and causation. We are taught that "correlation does not imply causation," but in non-linear dynamics, the problem is even deeper: variables can be causally linked while showing zero correlation, or they can appear highly correlated without any physical connection at all.

To solve this, researchers have turned to State-Space Reconstruction (SSR) and Convergent Cross Mapping (CCM). These methods move beyond looking at how two variables move together in time, instead looking at how they are embedded within the same underlying mathematical structure.

The Intuition: Nature Leaves a Paper Trail

The core idea rests on a principle from dynamical systems theory: if a variable $X$ influences a variable $Y$, then information about the state of $X$ must be encoded within the historical trajectory of $Y$. As Huanfei Ma et al. explain, in non-linear systems, variables are "non-separable." The influence of a driver is permanently recorded in the time series of the variable it affects.

Imagine a system as a high-dimensional shape called an attractor. While we might only observe a single variable (like temperature or a single neuron's activity), that variable is a "shadow" of the full system. According to Takens’ Theorem, we can reconstruct a version of the original attractor—a shadow manifold—using only time-delayed versions of that single observed variable.

Jakob Runge et al. and Yi Ding et al. describe how this allows us to test for causality. If $X$ causes $Y$, the shadow manifold reconstructed from $Y$ should be able to "predict" or recover the states of $X$. If we can look at the neighborhood of a point on $Y$'s manifold and use it to accurately estimate the corresponding point on $X$'s manifold, we have found a causal link.

Why it is "Convergent"

The "Convergent" in Convergent Cross Mapping refers to how the method behaves as we gather more data. In a truly causal relationship, as the length of the time series increases, the reconstructed shadow manifold becomes denser and more detailed. Consequently, the accuracy of the cross-mapping prediction should converge to a high value.

This distinguishes true causality from "mirage" correlations. As Isura Manchanayaka et al. demonstrate in the context of social media coordination, this convergence allows researchers to identify directed "influence flows" between agents that simple content similarity would miss. Similarly, Fang Guo et al. used this convergence to identify ozone as a hidden driver of influenza activity, even when linear correlations were weak.

Detecting Influence in the Real World

Because these methods are model-free—meaning they don't require us to write down the equations of the system beforehand—they are uniquely suited for "messy" real-world data:

  • Neuroscience: Satohiro Tajima et al. used cross-embedding to show how dynamical complexity accumulates as information flows from sensory regions to the frontoparietal "workspace" in the brain. André Fonseca et al. applied CCM to identify neural signatures of driver fatigue, mapping the effective connectivity between brain regions.
  • Earth Systems: Kristian Agasøster Haaga et al. re-evaluated the "pacemakers" of the ice ages, discovering that solar forcing is a spatially distributed system rather than a simple global correlation.
  • Biology: Bowen Yang developed the CAT Bridge toolkit, which uses CCM to identify biosynthetic pathways in multiomics data that traditional methods overlook.

Interestingly, Junjie Jiang et al. argue that noise, usually seen as a nuisance, can actually make these directed influences more detectable by helping to explore the system's state space more thoroughly.

Modern Frontiers: Deep Learning and Tangent Spaces

Recent work has sought to bring the power of deep learning to state-space reconstruction. Edward De Brouwer et al. introduced Latent CCM, which uses Neural ODEs to handle the sporadic, irregularly sampled data common in medicine. Matthew James Vowels et al. proposed Neural Shadow-Mapping, which can discover causal links directly from raw video data by learning the mapping between manifolds in an unsupervised way.

Finally, the field is moving toward even more robust geometric measures. Kurt Butler et al. suggest looking at tangent spaces—evaluating the alignment of "velocity vectors" across manifolds. By checking if the direction of change in one manifold consistently maps to the direction of change in another, researchers can identify causal coupling with unprecedented precision.

Key Foundational Work

Directed dynamical influence is more detectable with noise (2016) — contributed 1 concept elements to this direction.
Untangling Brain-Wide Dynamics in Consciousness by Cross-Embedding (2015) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_convergent_cross_mapping_43106c7c,
  title = {convergent cross mapping},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/convergent-cross-mapping/rdr_43106c7c21a745a1a49c01574ab71431},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_convergent_cross_mapping_43106c7c_v1,
  title = {convergent cross mapping},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/convergent-cross-mapping/rdr_43106c7c21a745a1a49c01574ab71431/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake