Technical Summary: Counterfactual Identifiability Beyond Global Monotonicity

1. Problem Statement

Structural Causal Models (SCMs) provide a unified semantics for interventions and counterfactuals. However, a fundamental challenge remains: counterfactual identifiability. Observational data alone is often insufficient to uniquely determine counterfactual outcomes unless strong assumptions are imposed on the mechanism class.

Historically, positive identifiability results have relied on global monotonicity (e.g., additive noise, post-nonlinear structures, or monotone triangular parametrizations). These assumptions ensure that the relationship between exogenous variables and endogenous outcomes preserves a fixed orientation.

The paper identifies a critical gap in embodied interaction systems (e.g., robotics): physical phenomena like contact, friction, and geometry often induce context-dependent sign reversals. In such scenarios, the same exogenous perturbation can produce opposite responses depending on the parent state (e.g., a push causing forward motion in one context and backward motion in another). Global monotonicity is too restrictive for these systems, yet removing structure entirely restores counterfactual ambiguity. The central question is: What structural constraints are necessary to preserve counterfactual identifiability once global monotonicity is abandoned?

2. Methodology and Theoretical Framework

2.1 The Proposed Model Class: NM-TM-SCM

The authors introduce Non-Monotone Triangular Structural Causal Models (NM-TM-SCM). This class retains the recursive triangular structure of standard SCMs but relaxes the global monotonicity constraint. Instead, it imposes two specific conditions:

1. Mechanism-wise Invertibility: For every mechanism $i$ and every fixed parent context $v_{<i}$, the mapping from the exogenous variable $u_i$ to the endogenous variable $v_i$ is a bijection. This ensures that factual abduction (recovering the exogenous state from observations) is unique given the parents.
2. Context-Independent Inverse Transport: For any two observation-equivalent models $M$ and $M'$, the composition of their inverse mechanisms, $(f'_i(v_{<i}, \cdot))^{-1} \circ f_i(v_{<i}, \cdot)$, must be independent of the parent context $v_{<i}$.

2.2 Theoretical Contributions

The paper establishes a theoretical chain linking these conditions to complete counterfactual identifiability:

• Lemma 1: Mechanism-wise invertibility implies that the global solution map $\Gamma_M: \Omega_U \to \Omega_V$ is a bijection.
• Theorem 1: Within the class of shared-order triangular SCMs, context-independent inverse transport is equivalent to exogenous isomorphism. Exogenous isomorphism means that two observationally equivalent models share the same latent worlds up to a coordinate change in the exogenous space that does not vary with context.
• Theorem 2: Exogenous isomorphism implies complete counterfactual identifiability. If two models are exogenously isomorphic, they must agree on all counterfactual queries.
• Proposition 1 (Sharpness): The authors provide a constructive counterexample showing that local invertibility alone is insufficient. If the inverse transport depends on the context (i.e., the orientation of the mechanism flips based on parents), observationally equivalent models can disagree on counterfactuals. This proves that context-independent inverse transport is a necessary condition, not just a convenient sufficient one.

2.3 Algorithm: CausalInverter

To instantiate these theoretical conditions, the authors propose CausalInverter, a structured learning model with three key components:

1. Triangular Invertible Layers: Models the structural equations as invertible flows.
2. Context-Dependent Orientation Gates: A discrete gate $s_{i,\theta}(v_{<i}) \in \{-1, +1\}$ that allows the mechanism to flip signs based on the parent context, enabling non-monotonicity.
3. Transport-Stability Regularization: A loss term ($L_{tr}$) that penalizes variations in the learned inverse transport across different parent contexts, enforcing the context-independent inverse transport condition.

The training objective combines negative log-likelihood with cycle consistency and transport stability regularizers.

3. Experimental Results

The authors evaluate CausalInverter on synthetic mechanisms and physical MuJoCo interaction tasks (Push and Door).

3.1 Synthetic Benchmarks

• Setup: The authors tested globally monotone, threshold-flip, and smooth-flip mechanisms across various dimensions and noise families.
• Findings:
  • In globally monotone settings, all structured baselines perform competitively.
  • In non-monotone settings (where parent-dependent orientation flips occur), standard baselines (ANM, monotone triangular, contextual bijective flows) fail to reduce counterfactual error.
  • CausalInverter significantly reduces Counterfactual Mean Squared Error (CF-MSE) and achieves high direction accuracy (~0.97) in recovering the orientation field.
  • A continuous "bridge" sweep varying non-monotonicity strength shows that the performance gain of CausalInverter over monotone baselines scales linearly with the degree of non-monotonicity.

3.2 Physical Interaction Tasks (MuJoCo)

The authors compare CausalInverter against a suite of state-based baselines (Linear Dynamics, MLP, Autoregressive, GRU, Transformer, Conditional Flow) under a balanced counterfactual sampling protocol.

• MuJoCo Door (Strong Non-Monotonicity):

  • This task exhibits strong context-dependent contact dynamics (NMS score $\approx 0.696$).
  • CausalInverter achieves perfect event-level counterfactual recovery (100% success on the formal split) and the lowest continuous counterfactual angle error among models that also achieve perfect event accuracy.
  • It demonstrates superior stability across multiple seeds compared to Transformer and Conditional Flow baselines.
  • Ablation studies confirm that removing the orientation gate causes the most severe degradation, validating the necessity of modeling context-dependent flips.

• MuJoCo Push (Weak Non-Monotonicity):

  • This task has a lower non-monotonicity score (NMS $\approx 0.175$).
  • Here, the structural bias of CausalInverter does not dominate. Stronger low-data state predictors (e.g., Linear Dynamics, MLP) remain competitive or superior.
  • The authors interpret this as a bias-variance boundary: the structural constraint is beneficial when non-monotonicity is strong enough to make counterfactual paths ambiguous, but can be over-regularizing when the task is nearly monotone.

4. Significance and Claims

The paper claims to identify a broader identifiable regime situated between globally monotone triangular models and unconstrained black-box world models.

• Theoretical Insight: The work demonstrates that global monotonicity is not the essential ingredient for counterfactual identifiability. Instead, mechanism-wise invertibility combined with context-independent inverse transport (equivalent to exogenous isomorphism) is the sufficient and sharp condition.
• Practical Impact: By relaxing the global monotonicity constraint, the proposed framework (NM-TM-SCM) and the CausalInverter model enable reliable counterfactual reasoning in physical systems where contact and geometry induce sign reversals.
• Boundary Awareness: The paper explicitly defines the limits of its contribution. The structural bias is most effective in regimes with strong branch-sensitive non-monotonicity (like the Door task). In weakly non-monotone regimes (like Push), flexible state predictors may outperform the structured approach, highlighting a tradeoff between causal identifiability constraints and task complexity.

The authors conclude that their approach provides a parameter-efficient route to event-level branch recovery in complex physical interactions, offering a structured alternative to purely predictive world models that lack causal guarantees.