Markov-Equivalent Conditional Independence Graphs on Path Spaces

Simon Kempf, Niki Kilbertus, Georg MantenarXiv 2026 · arXiv

TL;DR

This paper introduces W-separation, a new graphical criterion for causal discovery in continuous-time systems that allows for separate conditioning on the past and future of stochastic processes. By resolving structural gaps in previous methods, the authors prove that every set of causal dependencies has a unique "greatest" representative graph, providing a solid mathematical foundation for learning causal structures in complex, cyclic dynamical systems with hidden variables.

What Problem Does This Paper Address?

Inferring cause-and-effect relationships in dynamical systems (like financial markets or biological networks) is notoriously difficult. Researchers typically face a frustrating trade-off:

  1. Informative but untestable: Methods like "local independence" capture the flow of time perfectly but are nearly impossible to test statistically for common processes like diffusions.
  2. Testable but coarse-grained: Methods that discretize time or use simple path-space tests often lose the "arrow of time" or fail to account for cycles and hidden (latent) variables.

A recent attempt called E-separation tried to bridge this gap but had a major flaw: it couldn't perfectly distinguish between nodes that are causally connected and those that just appear to be due to hidden confounding. This paper introduces W-separation to fix these structural "blind spots" and provide a rigorous way to identify the most informative causal graph possible from observational data.

Methods

The authors propose a quaternary independence model. Instead of the standard "A is independent of B given C," they use: $$A \perp\perp_W B \mid C, D$$ where $C$ represents conditioning on the past of a set of processes, and $D$ represents conditioning on the future of another set.

To make this work, they use a lifted dependency graph $\tilde{\mathcal{G}}$. In this graph, every process $X$ is split into two nodes: $X_0$ (its past history) and $X_1$ (its future evolution). W-separation then applies standard graphical separation rules (like d-separation) to this expanded graph.

Figure 1: Comparison of E-separation and W-separation. Panel (d) shows how W-separation correctly identifies independence by conditioning independently on past and future nodes, whereas E-separation (panels a, b) fails to block certain paths.

Datasets

Not discussed in the paper. This is a theoretical work focused on establishing graphical and probabilistic foundations.

Key Insights and Ideas

  • The Power of Future Conditioning: The "aha moment" is that to truly block non-causal paths in continuous time, you sometimes need to condition on the future of a mediator. This sounds counter-intuitive but is mathematically necessary to handle latent confounding in cyclic systems.
  • Inducing Walk Equivalence: The authors prove that W-separation perfectly characterizes "a-inducing walks." In plain English: if two processes cannot be separated by any combination of past/future conditioning, there must be a specific causal path (or hidden common cause) connecting them.
  • The Unique Greatest Element: In causal discovery, many different graphs can often explain the same data (an "equivalence class"). The authors prove that for W-separation, every equivalence class has a unique greatest element—a single graph that contains all possible edges allowed by the data. This gives machine learning algorithms a clear, single target to aim for.

Key Results

  • Global Markov Property: The authors prove that W-separation is "sound" for a broad class of Stochastic Differential Equations (SDEs). If the graph says two processes are W-separated, they are guaranteed to be conditionally independent in the actual data.
  • Full Characterization of Equivalence: They provide necessary and sufficient conditions for when an edge can be added to a graph without changing its independence model.
  • Identifiability: Because of the unique greatest element property, the maximal causal structure is structurally identifiable from the W-independence model alone.

Limitations and Future Work

  • Computational Complexity: While the theory is robust, the search space for the optimal conditioning sets $(C, D)$ is combinatorially large. Developing efficient search algorithms is a key next step.
  • Statistical Power: While W-separation is "testable" in theory using path-space kernels, the practical sensitivity of these tests to sample size and noise in continuous-time data remains to be fully explored in empirical settings.
  • Algorithm Development: The paper provides the "rules" for causal discovery but leaves the implementation of a concrete discovery algorithm (like a modified PC or FCI algorithm) for future work.

Summary updated

Abstract

Causal discovery for continuous-time stochastic processes with latent confounding typically faces a trade-off between (i) approaches that are highly informative about the causal structure but are difficult to test statistically, and (ii) approaches that admit practical tests from data but yield more coarse-grained conclusions, potentially not leveraging the direction of time. We introduce W-separation, an asymmetric graphical criterion for directed mixed graphs defined as a quaternary relation that allows distinct conditioning on past and future components. This added flexibility yields a highly informative graphical independence model while preserving testability. W-separation satisfies a global Markov property for a broad class of stochastic differential equations with latent confounding, linking it to conditional independence relations among coordinate processes that are testable from data. A central result of this work is the full characterization of W-Markov equivalence classes: we show that every equivalence class admits a unique greatest element that serves as a canonical representative and a principled target for learning algorithms. In addition, we provide necessary and sufficient conditions under which directed or bidirected edges can be added without altering the induced independence model.
BibTeX
@misc{kempf2026markov,
  author = {Simon Kempf and Niki Kilbertus and Georg Manten},
  title = {{Markov-Equivalent Conditional Independence Graphs on Path Spaces}},
  year = {2026},
  eprint = {2611.20877},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.20877},
}
Figure