Approximate Causal Abstractions

Sander Beckers, Frederick Eberhardt, Joseph Y. HalpernUAI 2019 · OpenReview

Abstract

Scientific models describe natural phenomena at different levels of abstraction. Abstract descriptions can provide the basis for interventions on the system and explanation of observed phenomena at a macro level. While previous formalisms required exact mappings between levels, this paper introduces a framework to quantify the "causal distance" between models. This allows for a formal definition of how well a simplified abstract model approximates a complex underlying system.

Introduction

In scientific practice, we rarely utilize the most fundamental level of description to explain macro-phenomena. Instead, we rely on abstractions—simplified models that capture essential causal relationships while discarding micro-level details. Previous research defined "exact causal abstraction," where a high-level model perfectly mirrors the causal behavior of a low-level model under all possible interventions. However, real-world abstractions, such as the Ideal Gas Law or climate indices like "El Niño," are almost never exact. They are useful approximations. This paper addresses the lack of a formal, quantitative framework to define when a high-level model is a "good enough" causal abstraction of a complex low-level system.

Theoretical Framework

The authors extend the Structural Causal Model (SCM) framework to define approximate abstractions. The core of this framework involves a mapping function $\tau$ that translates low-level micro-variables to high-level macro-variables.

A central contribution is the introduction of the $\tau$-$\alpha$ Approximate Abstraction. A high-level model $M_H$ is considered an $\alpha$-approximate abstraction of a low-level model $M_L$ if the discrepancy between their predictions, across all possible interventions and contexts, is bounded by $\alpha$. To handle probabilistic models, the authors introduce "intervention distributions," which account for the fact that a single high-level intervention (e.g., "increase temperature") can be realized by multiple low-level configurations.

Quantifying the causal distance between high-level and low-level models through intervention distributions.

The framework introduces a causal distance metric, $d_{max}$, which measures the worst-case discrepancy. This metric is flexible, allowing researchers to choose an underlying distance measure ($d_V$) that reflects the priorities of their specific domain.

Case Study: El Niño Southern Oscillation

To illustrate the framework, the authors use a running example based on climate data. The El Niño Southern Oscillation (ENSO) is a complex phenomenon involving millions of data points for wind speed and sea-surface temperature (SST) across the Pacific.

Causal mapping of high-dimensional climate data (wind $\vec{W}$ and temperature $\vec{T}$) to a low-dimensional abstract model ($C$ and $E$).

In this example, the low-level variables (wind $\vec{W}$ and temperature $\vec{T}$) are mapped via $\tau$ to high-level variables ($C$ and $E$). The abstract model captures the essential causal link between central Pacific wind patterns and eastern Pacific temperature anomalies, providing a tractable model for climate prediction despite being an approximation of the underlying atmospheric physics.

Key Results and Insights

  • Intervention-Centric Accuracy: The validity of an abstraction is determined not just by its descriptive accuracy of the current state, but by its ability to predict the outcomes of system manipulations.
  • Handling Ambiguity: By utilizing intervention distributions, the framework provides a rigorous way to manage the inherent underspecification of high-level interventions at the micro-level.
  • Decomposition and Complexity: The authors explored whether approximate abstractions can always be decomposed into an "Exact Abstraction + Approximation." They proved this is true for "constructive" abstractions (where variables are partitioned into disjoint sets) and that verification in these cases is computationally efficient (polynomial-time), whereas the general case is NP-hard.
  • Probabilistic Extension: The framework successfully incorporates uncertainty in both environmental contexts and the implementation of interventions, making it applicable to stochastic real-world systems.

Conclusion and Future Directions

This work provides a rigorous mathematical foundation for evaluating simplified scientific models. By formalizing the "causal distance," it bridges the gap between idealized exact abstractions and the approximate models used in practice.

Future research directions include the development of standardized metrics for specific scientific domains, extending the framework to handle models with feedback loops (cyclic models), and creating algorithms to automatically discover optimal $\alpha$-approximate abstractions from raw observational and experimental data.

Summary updated

Abstract

Scientific models describe natural phenomena at different levels of abstraction. Abstract descriptions can provide the basis for interventions on the system and explanation of observed phenomena at...
BibTeX
@inproceedings{beckers2019approximate,
  author = {Sander Beckers and Frederick Eberhardt and Joseph Y. Halpern},
  title = {{Approximate Causal Abstractions}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2019},
  url = {https://openreview.net/forum?id=irWvRPrTRA6},
}
Figure