Causal Discovery for Multivariate Groups and Vector Variables

This direction is merged into another direction: non-linear causal discovery. Its last version stays readable and citable here.

Version 1 of 19 papers18 conceptsfrozen 2026-10-04

In fields like neuroscience, climate science, and genomics, we rarely measure a single "thing." Instead, we measure groups of related variables: the activity of thousands of neurons in a brain region, temperature readings across a geographic grid, or the expression of hundreds of genes.

To make sense of this complexity, the standard practice is to simplify. We average the voxels in a brain region or the temperatures in a grid to create a single representative scalar. However, a growing body of research suggests that this simplification is a "causal signal killer." By aggregating data before analyzing it, we inadvertently erase the very evidence needed to distinguish cause from effect.

The Gaussian Trap and the Central Limit Theorem

Many modern causal discovery algorithms rely on non-Gaussianity. In a linear world, if you see a relationship between $X$ and $Y$, you can't tell if $X \to Y$ or $Y \to X$ just by looking at correlations. But if the noise is non-Gaussian, the "residual" (the part of the effect that the cause doesn't explain) will only be independent of the cause in the correct direction.

The problem is that aggregation—whether spatial or temporal—is a natural "Gaussianizer." As Yasuhiro Sogawa, Aapo Hyvarinen, and colleagues noted in their work on the EggFinder algorithm, the Central Limit Theorem dictates that as signals propagate and mix, they converge toward a Gaussian distribution.

This "statistical convergence" is particularly destructive in time-series data. Shunxing Fan, Kun Zhang, and others have shown that temporal aggregation erases the unique distributional signatures required by functional causal models. As the aggregation window grows, the non-Gaussian signals that distinguish cause from effect effectively vanish, leaving behind a symmetric Gaussian blur where directionality is unrecoverable.

Why Averaging Fails: Cancellation and Dilution

Beyond changing the shape of the distribution, aggregation causes physical information loss. When we reduce a high-dimensional group to a single average, we risk signal cancellation.

Imagine a brain region where half the neurons excite a downstream region and the other half inhibit it. If you average the activity of the first region, these opposing effects may cancel out to zero. To an observer looking only at the average, it appears as if no causal relationship exists at all. Jonas Wahl, Jakob Runge, and Urmi Ninad argue in their foundational work that this "information dilution" makes it impossible to detect granular causal links.

Furthermore, the transition from "micro" variables (individual neurons) to "macro" variables (brain regions) breaks the Causal Faithfulness assumption. Faithfulness assumes that if two variables are statistically independent, they are not causally linked. But as Wahl et al. demonstrate, macro-level variables often exhibit dependencies that don't exist at the micro-level, or hide dependencies that do, leading to "spurious" causal maps.

Vector-Valued Causal Inference

The solution is to stop treating groups as scalars and start treating them as vectors. Instead of asking "Does the average of Group A cause the average of Group B?", we ask "Does the multivariate state of Group A cause the multivariate state of Group B?"

Wahl, Runge, and Ninad introduced a framework called Vector Causal Inference (VecCI). Rather than collapsing the data, their algorithms (like 2G-VecCI) analyze the "internal messiness" or edge density within a group. By looking at how the internal structure of Group B changes when we account for Group A, they can determine the direction of influence without losing granular detail.

This approach has been extended to handle even more complex scenarios:

  • Adaptive Aggregation: The Adag wrapper, developed by Urmi Ninad and colleagues, automatically finds the "sweet spot" of aggregation. It ensures that any simplification of the data remains mathematically consistent with the underlying micro-level interactions.
  • Context-Specific Links: Systems often change their causal "rules" based on internal states. The PC-AC algorithm by Wiebke Günther and Andreas Gerhardus avoids the trap of pooling all data together, instead adaptively identifying different operating regimes to capture links that only appear in specific contexts.

Summary: From Scalars to Systems

The shift toward vector-based causal inference represents a move away from "black box" aggregation toward a more principled treatment of complex systems. By acknowledging that groups of variables contain structured information that cannot be reduced to a single mean, researchers can bypass the limitations of the Central Limit Theorem and recover causal structures that were previously invisible. As Wahl et al. emphasize, the goal is to build macro-level models that are not just convenient summaries, but accurate reflections of the underlying micro-level dynamics.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_causal_discovery_for_multivariate_groups_8d41b33d,
  title = {Causal Discovery for Multivariate Groups and Vector Variables},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-discovery-for-multivariate-groups-and-vector-variables/rdr_8d41b33d8ee04720a45873f14633d470},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_causal_discovery_for_multivariate_groups_8d41b33d_v1,
  title = {Causal Discovery for Multivariate Groups and Vector Variables},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-discovery-for-multivariate-groups-and-vector-variables/rdr_8d41b33d8ee04720a45873f14633d470/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}