Bijective Causal Models for Counterfactual Identifiability

18 concepts14 papers1 proposal

Imagine you are looking at a historical data point—say, an individual who was denied a loan. A counterfactual question asks: "What would the outcome have been if this specific person's income had been higher?"

Answering this requires traversing Judea Pearl’s Causal Hierarchy to its highest tier (Level 3). According to structural causal theory, evaluating a counterfactual follows a strict three-step recipe:

  1. Abduction: Use observed data to infer the unobserved background conditions (exogenous noise, $U$) that drove the event.
  2. Action: Modify the system by intervening on a variable (e.g., setting income to a higher value).
  3. Prediction: Simulate the modified system to calculate the new outcome.

The bottleneck is almost always Abduction. In a standard Structural Causal Model (SCM), an observed variable $X$ is generated by a function of its causal parents and some exogenous noise: $X = f(PA_X, U)$. If $f$ is a complex, many-to-one function, infinitely many different noise configurations $U$ could produce the exact same observation $X$. Without knowing which $U$ actually occurred, the counterfactual outcome is mathematically unidentifiable—meaning observational data cannot uniquely determine the answer.

A rapidly growing body of research solves this fundamental dilemma by enforcing a structural constraint: bijectivity. By ensuring a one-to-one, invertible mapping between the observed data and the latent exogenous noise, these frameworks make the abduction step entirely deterministic, unlocking complete counterfactual identifiability.


The Core Intuition: Why Bijectivity Unlocks Abduction

To understand why bijectivity is so powerful, consider the geometry of the latent space. If the function $f$ mapping exogenous noise $U$ to an effect $X$ is smooth and bijective for any fixed value of its causal parents, it possesses a well-defined inverse, $f^{-1}$.

As explained by Arash Nasr-Esfahany et al. in Counterfactual Identifiability of Bijective Causal Models, this Bijective Generation Mechanism (BGM) guarantees that the true, unobserved background noise can be uniquely recovered from what we observe. There is no ambiguity or loss of information during the abduction step.

Standard SCM (Many-to-One)         Bijective SCM (One-to-One)
   Latent       Observed              Latent       Observed
   Noise         Data                 Noise         Data
   [ U₁ ] \                              
   [ U₂ ] ———>  [ X ]                 [ U ]  <======>  [ X ]
   [ U₃ ] /                           (Unique Inverse)
(Ambiguous Abduction)

This structural assumption is surprisingly flexible. Kasra Jalaldoust et al. demonstrate in Multi-Domain Causal Discovery in Bijective Causal Models that bijective causal mechanisms generalize several classic functional forms—such as additive noise models and location-scale models—while accommodating highly complex, non-linear interactions between causes and noise.

Furthermore, we do not even need to recover the exact physical latent variables to achieve identifiability. Yikang Chen et al. show in Exogenous Isomorphism for Counterfactual Identifiability that as long as two models share an isomorphic latent noise space (a bijective mapping between their respective latents), they will yield identical counterfactual predictions. This insight allows researchers to build Neural Triangular Monotonic SCMs that guarantee counterfactual consistency even for high-dimensional, vector-valued data.


Scaling to High Dimensions with Deep Generative Models

While the mathematics of bijectivity are elegant, parameterizing these invertible functions for high-dimensional, real-world data requires deep learning. A comprehensive taxonomy by Audrey Poinsot et al. in Learning Structural Causal Models through Deep Generative Models: Methods, Guarantees, and Challenges highlights how deep generative architectures can be constrained to preserve these identifiability guarantees.

Deterministic Diffusion as a Causal Bridge

One of the most powerful realizations of this concept leverages diffusion models. While standard diffusion processes rely on stochastic sampling, they can be formulated deterministically using the Probability Flow Ordinary Differential Equation (ODE).

Patrick Chao et al. demonstrate in Modeling Causal Mechanisms with Diffusion Models for Interventional and Counterfactual Queries that this deterministic ODE acts as a perfect bi-directional bridge between complex data distributions and standard Gaussian noise. Integrating the ODE forward in time maps an observation to a unique noise vector (Abduction), while integrating it backward generates the counterfactual outcome (Prediction).

Observation (X)  ================== Forward ODE (Abduction) ==================>  Unique Noise (U)
                                                                                       ||
Counterfactual (X*) <=========== Backward ODE (Action + Prediction) ===========  Intervention / Fix U

This deterministic latent mapping is highly effective for complex environments. For instance, Xinwen Liu et al. develop the Partially Functional Dynamic Backdoor Diffusion-based Causal Model, which embeds backdoor adjustments directly into the implicit diffusion process. This allows the model to maintain consistency between observed facts and counterfactual simulations even across space, time, and hidden confounders.

This deep connection between generative trajectories and causal variables is formalized by Amir Mohammad Karimi Mamaghan et al. in Diffusion Based Causal Representation Learning (and expanded in Diffusion-Based Causal Representation Learning). The authors prove that the infinite-dimensional latent trajectories derived from diffusion processes are empirically and theoretically interchangeable with the exogenous noise variables of an SCM, providing a robust foundation for causal discovery.


Applications: Fairness, Recourse, and Robustness

Once a model can uniquely isolate exogenous noise via bijective mappings, it can reliably answer individual-level "what-if" questions across several domains.

Counterfactual Fairness and Invariance

In algorithmic decision-making, fairness often requires that a model's prediction remains unchanged had an individual belonged to a different demographic group. Yuchen Ma et al. implement this in Counterfactual Fairness for Predictions using Generative Adversarial Networks by using bijective generation mechanisms within a GAN framework to uniquely recover counterfactual distributions without ambiguity.

Similarly, Francesco Quinzan et al. show in Learning Counterfactually Invariant Predictors that assuming injective structural equations allows researchers to transform complex counterfactual queries into verifiable statistical independence tests, enabling the training of predictors that are robust to environmental shifts.

Algorithmic Recourse and Explanations

When an AI system outputs an unfavorable decision, algorithmic recourse aims to provide actionable recommendations for the user. Pouria Fatemi et al. introduce a unified framework in A New Approach to Backtracking Counterfactual Explanations: A Unified Causal Framework for Efficient Model Interpretability that maps observed features back to exogenous noise. Optimization is then performed directly within this latent space, ensuring that suggested changes are physically and logically plausible.

For sequential decision-making, Stratis Tsirtsis et al. demonstrate in Finding Counterfactually Optimal Action Sequences in Continuous State Spaces that fixing these uniquely identified noise terms allows an agent to simulate alternative action sequences while keeping the underlying environmental stochasticity perfectly constant.

Deconfounding and Self-Supervision

Bijectivity also provides strong mathematical guarantees when dealing with unobserved variables. In proximal causal inference, Alejandro Almodóvar et al. prove in DeCaFlow: A deconfounding causal generative model that individual-level counterfactuals are fully identifiable from proxies if the underlying generative mapping is bijective.

Finally, this principle enables unsupervised structure learning; Xiaoyang Wang et al. show in Identifying Coarse-grained Independent Causal Mechanisms with Self-supervision that combining an invertible mapping with a structured prior provides a mathematical guarantee that a model can disentangle independent causal mechanisms from raw, unstructured data without requiring any auxiliary labels.

By framing deep generative models as invertible bridges to latent noise, this research direction transforms counterfactual reasoning from an unidentifiable philosophical problem into an exact, computable science.

Key Foundational Work

Counterfactual Identifiability of Bijective Causal Models (2023) — contributed 2 concept elements to this direction.
Counterfactual Fairness for Predictions using Generative Adversarial Networks (2023) — contributed 1 concept elements to this direction.
Identifying Coarse-grained Independent Causal Mechanisms with Self-supervision (2022) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_deterministic_latent_mapping_3a64a2cf,
  title = {deterministic latent mapping},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/deterministic-latent-mapping/rdr_3a64a2cf3d564eceb1a0f4b9834a5ab6},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_deterministic_latent_mapping_3a64a2cf_v1,
  title = {deterministic latent mapping},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/deterministic-latent-mapping/rdr_3a64a2cf3d564eceb1a0f4b9834a5ab6/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake