Bow-free Identifiability in Causal Models

• Bow-free identifiability is a structural property that ensures unique and stable parameter recovery by excluding overlapping directed and bidirected edges.
• In models like LSEMs and BAPs, the absence of bows leads to both generic and robust identifiability, meaning small data perturbations yield minimal changes in estimates.
• Algorithmic approaches and empirical validations demonstrate that enforcing bow-free constraints enhances structure learning and reliable causal effect estimation in practical settings.

Bow-free identifiability is a structural property in models of causality and latent variable analysis, denoting generic or robust recoverability of model parameters when certain graphical constraints—specifically, the exclusion of "bows"—are satisfied. In both linear structural equation models (LSEMs) and mixed graph models such as bow-free acyclic path diagrams (BAPs), the absence of bows ensures that identifiability holds almost everywhere in parameter space, and, critically, can often be strengthened to robust identifiability: small perturbations in observed statistics yield only small changes in estimated parameters. Recent advances have also generalized bow-free identifiability notions to deep generative models, showing that in certain "bow-free" settings, unsupervised recovery of latent structure is theoretically guaranteed.

1. Bow-free Models and Their Formalization

A bow in a mixed graph occurs when the same pair of vertices is connected by both a directed edge (representing direct causal influence) and a bidirected edge (representing shared unmeasured confounding). A model is bow-free if no such pair exists: i.e., for all $i \ne j$, at most one of $i \to j$ or $i \leftrightarrow j$ is present. In linear-Gaussian SEMs, the underlying mixed graph $G=(V, E, F)$ encodes directed edges $E$ for acyclic causal relationships and bidirected edges $F$ for covariance structure representing latent confounding (Sankararaman et al., 2020, Nowzohour et al., 2015).

The linear structure can be written as: $$X = \Lambda^T X + \varepsilon,\qquad \varepsilon\sim \mathcal{N}(0, \Omega)$$ with observed covariance

$$\Sigma = (I - \Lambda)^{-T} \Omega (I - \Lambda)^{-1}$$

where $\Lambda$ has nonzero entries for $u\to v\in E$, and $i \to j$0 for $i \to j$1. The bow-free criterion imposes that if $i \to j$2 then $i \to j$3, and vice versa.

Bow-free acyclic path diagrams (BAPs) are a subclass characterized by DAG structure for directed edges and the absence of bows, thus generalizing DAG models to permit partial hidden confounding within strict graphical constraints (Nowzohour et al., 2015).

2. Generic and Robust Identifiability in Bow-free Models

Generic Identifiability

A bow-free LSEM or BAP is generically identifiable if for almost all permitted parameters, the forward map from parameters $i \to j$4 to the observed covariance $i \to j$5 is locally one-to-one. That is, for almost all parameter values, the Jacobian of this map has full rank, ensuring the existence of a unique (up to measure-zero sets) inverse mapping from $i \to j$6 to $i \to j$7. In the specific context of BAPs, the non-identifiable set has Lebesgue measure zero in the parameter space (Nowzohour et al., 2015).

Robust Identifiability

Going beyond genericity, robust identifiability requires that small perturbations in $i \to j$8 lead only to small perturbations in $i \to j$9, quantified by a relative $i \leftrightarrow j$0 condition number: $i \leftrightarrow j$1 where $i \leftrightarrow j$2 denotes entrywise maximal relative deviations. Bow-free models, under mild spectral (well-conditioning) and diagonal-dominance assumptions, admit sufficient conditions ensuring that this condition number is polynomially bounded in $i \leftrightarrow j$3, thus guaranteeing stable parameter recovery under finite-sample or estimation noise (Sankararaman et al., 2020).

3. Identifiability Characterizations and Equivalence Classes

Distributional equivalence in bow-free models is fully characterized by the skeleton (underlying undirected structure) and collider-triples (nodes with two directed or bidirected edges into the same node, regardless of adjacency between sources). Two BAPs with identical skeletons and collider-triples generate exactly the same class of normalized covariance matrices. As a result, model parameters and certain causal effects are only point-identified within a distributional equivalence class (Nowzohour et al., 2015).

The identifiability of specific causal effects (e.g., entries of the matrix $i \leftrightarrow j$4 in SEMs) depends on the full equivalence class. Effects that are invariant across all equivalent models are point-identified, while others are bounded by the range they take within the class.

Notion	Condition	Outcome
Generic identifiability	Full-rank Jacobian except null set	Unique recovery except for a measure-zero set
Robust identifiability	Spectral + diagonal-dominance assumptions	Parameter changes scale with covariance errors
Distributional equivalence	Identical skeleton and collider-triples	Indistinguishable model class via $i \leftrightarrow j$5

4. Algorithmic Approaches: Recovery and Structure Learning

Parameter recovery in bow-free SEMs leverages the half-trek RICF algorithm and maximum-likelihood estimation. The half-trek method solves for parameters layer by layer using small well-conditioned linear systems, provided bow-freeness and associated assumptions hold (Sankararaman et al., 2020, Nowzohour et al., 2015).

Structure learning for BAPs employs greedy, score-based algorithms that optimize a penalized likelihood over the restricted class of bow-free graphs. Local moves (edge addition/deletion, orientation flips) never introduce bows or cycles. Empirical equivalence classes can be enumerated using neighborhood search, guided by skeleton and collider-triple invariants, enabling identification of effects that are robust to model ambiguity (Nowzohour et al., 2015).

5. Bow-free Identifiability in Deep Generative Models

Recent theory situates bow-free identifiability within deep generative modeling. Consider

$i \leftrightarrow j$6

with a piecewise-affine (ReLU or leaky-ReLU) decoder $i \leftrightarrow j$7 and a mixture-of-Gaussians prior. The "bow-free" condition is operationalized by not observing or conditioning on any auxiliary side variable (as is needed to "break the bow" in previous nonlinear-ICA or iVAE frameworks).

The main results (Kivva et al., 2022) establish a four-level hierarchy:

• Level 1: $i \leftrightarrow j$8 identifiable up to affine transforms, under minimal prior and decoder assumptions.
• Level 2: $i \leftrightarrow j$9 identifiable up to permutation, rescaling, translation with stronger (diagonal+generic ratio) prior.
• Level 3: Discrete latent $G=(V, E, F)$0 (and law) identifiable up to permutation (if present).
• Level 4: Globally injective decoder is recoverable up to the same affine transformation.

No auxiliary labels or contrastive pairs are needed: standard VAE-style models with mixtures in the latent space and suitable decoders guarantee self-identification of the latent distribution and decoder function, modulo inherent linear indeterminacies.

6. Empirical Validation and Practical Implications

Conditional on well-behaved parameter regimes (e.g., bounded edge weights, diagonal dominance, input matrix conditioning), empirical simulations and real data applications (e.g., gene expression networks) confirm that robust identifiability is realized for sparse bow-free models (Sankararaman et al., 2020). For deep generative models, synthetic and real-world data experiments corroborate that unsupervised learning under bow-free conditions achieves the theoretically promised identifiability across the latent space and decoder (Kivva et al., 2022).

In practical structure learning, the enumeration of (nearly) equivalent bow-free models enables principled computation of lower bounds for causal effects when point identification is unavailable (Nowzohour et al., 2015).

A plausible implication is that bow-free identifiability results both increase the reliability of parameter inference in graphical causal analysis and expand the regime of unsupervised generative modeling where model-based recovery is theoretically guaranteed.

7. Ongoing Directions and Further Considerations

Research directions include relaxing diagonal-dominance conditions in robust identifiability, extending theory to semi-Markovian or partially observed models, and combining robustness criteria with model-misspecification. For deep generative models, extensions to broader classes of priors and decoder functions are a natural continuation. The generalization of bow-free identifiability theory across statistical, causal, and generative modeling frameworks suggests a unified principle for structural recovery absent conflicting confounding and direct effects.

Key contributions on bow-free identifiability and its theoretical, algorithmic, and empirical foundations are documented in (Sankararaman et al., 2020, Nowzohour et al., 2015), and (Kivva et al., 2022).