Bidirected Causal Connections

• Bidirected causal connections are symmetric relationships indicating mutual influence through latent confounders or cyclic feedback in both classical and quantum models.
• They are modeled using mixed graphs and structural equations to capture noise covariances and identify causal parameters via decomposability conditions.
• Applications range from constraint-based causal discovery in high-dimensional data to guiding entanglement summoning and robust inference in quantum networks.

A bidirected causal connection is a structural and graphical feature within models of causal inference, quantum information theory, and network theory that encodes symmetric influence or mutual information flow between two variables or nodes. Bidirected edges, denoted $i \leftrightarrow j$, represent the possibility of bilateral causal association, commonly interpreted as arising from shared latent confounders, feedback mechanisms, bidirectional communication channels, or symmetric operational constraints. Their rigorous treatment underpins advances in constraint-based causal discovery, graphical structural equation modeling, additive noise models, cyclic processes, and resource theory for quantum information.

1. Formal Definitions and Causal Semantics

Bidirected edges are integral to mixed graphs, structural equation models, $\sigma$-connection graphs, and process-matrix frameworks:

• Graphical Representation: In a mixed graph $G = (V, D, B)$, directed edges $i \to j$ encode direct causal effects, while bidirected edges $i \leftrightarrow j \in B$ indicate symmetric influence or unobserved common causes between $i$ and $j$. No bidirected self-loops are permitted (Fox et al., 2013).
• Causal Interpretation: The standard reading is that $i \leftrightarrow j$ encodes at least one unobserved common parent (latent confounder) $U$, with $U \to i$ and $U \to j$ (Forré et al., 2018).
• Quantum Networks: In quantum process networks, a bidirected causal connection between nodes $i$ and $j$ indicates that both parties may simultaneously send and receive quantum or classical information in a given communication round (Bozanic et al., 21 Jan 2026).
• Structural Equations: In parametric and semiparametric SCMs, bidirected edges correspond to nonzero covariance between error terms, typically induced via latent variables (Zander et al., 2022, Pham et al., 11 Feb 2025).

2. Representation in Structural Causal Models and Mixed Graphs

Structural equation models and mixed graphs represent bidirected causal connections as confounding correlations or feedback loops.

• Gaussian Mixed Graphs: Each bidirected edge $u \leftrightarrow v$ assigns a noise covariance $\omega_{uv}$ in $\Omega$, the error covariance matrix. The joint covariance is given by $\Sigma = (I-\Lambda)^{-T} \Omega (I-\Lambda)^{-1}$, with $\Lambda$ capturing directed effects (Fox et al., 2013, Zander et al., 2022).
• Exact Causal Interpretation: A mixed graph admits a strict causal interpretation (can be reproduced exactly by a hidden-variable DAG) if and only if the bidirected part is decomposable (chordal), i.e., every cycle of length $\geq 4$ has a chord. For non-decomposable bidirected parts, Markov equivalence fails (Fox et al., 2013).

Representation Model	Bidirected Causal Role	Key Condition
Mixed Graph Gaussian SEM	Latent confounders, error covariance	Decomposability of B
$\sigma$-Connection Graph	Latent parent, cyclic relationships	σ-separation closure
ADMG (Acyclic Dir. Mixed)	Unobserved paths, backdoors, feedback	CI and additive noise tests

3. Bidirected Edges in Constraint-Based Causal Discovery

Bidirected edges significantly impact the identifiability and inference mechanisms of causal structure learning.

• σ-Connection Graphs: Forré and Mooij’s framework allows bidirected edges, undirected, and directed edges; σ-separation is a generalization of d-separation for cyclic, non-linear, and confounded models (Forré et al., 2018).
• Marginalization and Conditioning Closure: σ-separation is closed under marginalization and conditioning, ensuring that the effect of latent confounders (bidirected edges) is precisely maintained (Forré et al., 2018).
• Algorithmic Implications: Causal discovery algorithms employ z-σ-open path blocking to test for conditional independence including latent confounders, yielding robust causal structure in the presence of bidirected (latent) edges (Forré et al., 2018).

4. Identification and Estimation via Bidirected Structure

Bidirected cycles and missing bidirected edges enable identification of causal parameters by algebraic means in linear and tree-shaped SEMs.

• Missing Cycle Equations: In tree-shaped models, the absence of a bidirected cycle induces a system of quadratic equations connecting observed covariances to directed-edge parameters. These equations may yield one or two candidate solutions, and intersecting multiple cycles typically identifies parameters uniquely (Zander et al., 2022).
• TreeID Algorithm: Sequential exploitation of missing bidirected cycles reduces reliance on doubly-exponential Gröbner basis computations, achieving polynomial time complexity for many real-world graphs (Zander et al., 2022).
• Instrumental Variables and Bi-directionality: In bidirectional models with invalid instruments, mode-clustering (plurality rule) and covariance heterogeneity criteria extract causal directionality even with bidirected feedback or symmetric confounding (Li et al., 2024).

5. Advanced Frameworks: Additive Noise, Quantum Processes, Resource Theory

Bidirected causal connections play foundational roles in advanced causal discovery and operational quantum information.

• Causal Additive Models (CAM-UV-X): Bidirected edges encode latent confounding or hidden causal paths. The CAM-UV-X algorithm exploits regression-set independence and conditional independence tests to identify and orient bidirected connections that are non-discernible by FCI or score-based methods (Pham et al., 11 Feb 2025).
• Quantum Resource Theory: In the process matrix formulation, bidirectional signalling capacity is captured via robustness monotones $R_s(W)$ and structural properties of process adapters. Maximal resourceful processes are those fully bidirectional combs ($W^{A\to B}$, $W^{B\to A}$), saturating tight dimensional bounds on $R_s$ (Milz et al., 2021).
• Entanglement Summoning Tasks: In quantum networks, a bidirected-only connection graph admits entanglement summoning if and only if the graph is partitionable into two bidirected cliques (equivalent to no odd cycles in the complement). This graphical condition precisely characterizes the achievable communication complexity for entanglement tasks (Bozanic et al., 21 Jan 2026).

6. Interpretational and Graph-Theoretic Criteria

The existence and meaning of bidirected causal connections depend on precise graph-theoretic and operational conditions.

• Chordality (Decomposability): Exact latent confounder interpretation for bidirected edges in a mixed graph is possible if and only if the undirected graph of bidirected edges is chordal/decomposable (Fox et al., 2013).
• Two-clique Partition: For bidirected quantum networks, the task of routing entanglement to requested sites is feasible if and only if the network splits into two fully bidirected cliques. Absence of odd cycles in the complement graph is necessary and sufficient (Bozanic et al., 21 Jan 2026).
• Feedback and Cyclic Order: Bidirected edges generalize to cyclic causal orders and feedback processes, as in models admitting bi-directional causal influence or open feedback loops (Li et al., 2024).
• Limitation and Failure Cases: Non-decomposable (e.g., chordless cycles of bidirected edges) undermine strict latent-variable interpretation and induce conditional independence patterns unreproducible by finite-dimensional linear-Gaussian models (Fox et al., 2013).

7. Applications and Experimental Implications

Bidirected causal connections have profound practical and theoretical significance.

• High-dimensional Genomics and Phenotypic Causality: In observational genetics, bidirected models with invalid IVs enable rigorous detection of causal directionality between traits under complex feedback and confounding (Li et al., 2024).
• Quantum Networking and Communication Complexity: Characterizing the summoning capacity of quantum networks with bidirected connections informs architecture and resource allocation for distributed quantum protocols (Bozanic et al., 21 Jan 2026).
• Constraint-based Causal Discovery in Non-linear and Cyclic Systems: σ-separation-based algorithms support reliable structure learning in domains where latent confounding and cycles are unavoidable (Forré et al., 2018).
• Operational Quantum Protocols: Resource-theoretic scoring for causal connection quantifies the maximal signalling capacity, guiding process engineering in indefinite-causal-order settings (Milz et al., 2021).

Bidirected causal connections thus constitute a core concept for modern causal inference, graphical modeling, and quantum information science, enabling technically rigorous treatment of symmetric association, latent confounding, feedback dynamics, and bi-directional signalling across diverse domains.