Optimal Transport Redefines Backtracking Counterfactual Inference

Summary:The paper formulates backtracking counterfactual inference in structural causal models as an optimal transport problem. A factual observation induces a posterior over exogenous variables, while a counterfactual query defines a constrained set of admissible exogenous configurations. Two complementary transports are introduced: a natural pointwise projection for direct use, and an independence-preserving formulation that keeps transported variables interpretable as genuine exogenous noise. For fi

---

Core Idea: Counterfactual Inference as Distribution Transport

In a structural causal model (SCM), observing a factual outcome $Y=y$ generally does not identify a unique value of the exogenous variables $U$. Instead, it induces a posterior:

$$p(U\mid Y=y)\propto p(y\mid U),p(U).$$

Counterfactual inference must therefore transport the posterior onto the set of exogenous configurations compatible with the desired query. The paper formalizes this as an optimal transport problem from $p(U\mid Y=y)$ to an admissible counterfactual set $\mathcal{C}_{cf}$.

This formulation makes three objectives explicit: minimizing transport cost, satisfying the counterfactual query, and preserving the semantic validity of the transported exogenous variables.

Two Complementary Frameworks

1. Pointwise Projection

The first framework maps each support point in the factual posterior to its nearest feasible counterfactual configuration:

$$T^*(u)=\arg\min_{u'\in\mathcal{C}_{cf}} d(u,u').$$

This is useful when the transport can be applied directly and the feasible set is easy to enumerate or query. However, independent pointwise projections may damage the original joint structure of the exogenous variables. Two variables that were correlated in the factual posterior can become artificially coupled after satisfying a query constraint.

2. Independence-Preserving Transport

The second framework operates on the joint distribution and seeks a transported distribution whose exogenous components remain mutually independent while satisfying the counterfactual query. This preserves the interpretation of each component as genuine exogenous noise in the same SCM rather than a latent variable created by coupling.

The independence requirement is semantically stronger, but it increases optimization difficulty. The resulting transport may incur greater cost or weaker query satisfaction than a simple projection.

Solvers and Constraint Handling

For finite posteriors, the paper proposes three solver strategies:

• Marginal solvers decompose the problem by variable or constraint and work best when the feasible structure is approximately separable.
• Penalty-based solvers add constraint violations to the transport objective. They are easy to implement but require careful tuning and may leave residual violations.
• Hard-constrained solvers restrict optimization to the feasible set. They provide auditable constraint satisfaction at the cost of higher computation and potentially larger distributional distortion.

The appropriate choice depends on the application. Hard constraints are preferable when violating a counterfactual condition is unacceptable, such as in high-stakes policy or medical simulations. Penalty methods may be more practical for sensitivity analysis or large-scale approximate sampling.

Engineering Trade-offs

The central engineering issue is that transport cost, query satisfaction, and exogenous independence cannot generally be maximized simultaneously:

1. Low transport cost keeps outputs close to the factual posterior but may provide only weak satisfaction of the counterfactual query.
2. Strong constraints improve counterfactual validity but can move samples far from the original posterior support.
3. Preserving independence supports SCM interpretability but may require additional marginal or joint distortion.

With finite posteriors, enumeration and re-optimization can become expensive as sample count, variable dimension, and constraint count grow. Penalty methods are easier to scale, yet their results depend on initialization, penalty weights, and iteration budgets. The experiments focus on non-bijective SCMs and expose geometric behavior; they do not establish scalability for high-dimensional industrial SCMs.

Interpreting the Evaluation

The synthetic experiments are most informative as a demonstration of unavoidable trade-offs rather than a universal ranking. A convincing evaluation should independently verify that the generated counterfactual actually satisfies the intended intervention or outcome condition. It should also report the cost metric, constraint violation rate, and a joint-distribution test for independence instead of relying only on marginal means and variances.

The current results do not demonstrate robustness under continuous high-dimensional SCMs, model misspecification, or observational noise. The work is therefore best viewed as a principled solver framework and evaluation baseline for counterfactual inference.

Practical Adoption Guidance

• Start with a pointwise projection on a low-dimensional, enumerable SCM to establish baseline cost and feasibility trade-offs.
• Add independence-preserving transport when downstream interpretation requires genuine exogenous-noise semantics.
• Evaluate hard and soft constraint modes separately rather than collapsing them into one satisfaction metric.
• Retain the factual posterior support and transport maps for auditing.
• Before scaling, test whether the problem admits separability, low-rank structure, or exploitable causal directions that can reduce optimization cost.

---

Source: ArXiv AI (cs.AI) (2026-11-09)

— END —

Tags: #Optimal Transport Follow #Structural Causal Models Follow #Counterfactual Inference Follow #Exogenous Independence Follow #Causal Inference Follow