When we reason about the world, we rarely operate at the level of fundamental physics. A macro-economist studies inflation and unemployment rather than individual financial transactions; a cognitive scientist analyzes beliefs and desires rather than individual neurons firing. In all these cases, we are constructing a simplified, high-level model (a macro-model) of a highly complex, low-level reality (a micro-model).
In causal inference, this process of simplification is formalized as causal abstraction. However, simplifying a model runs a severe risk: if we group variables together incorrectly, our high-level model might make completely wrong predictions about what happens when we intervene on the system.
The core insight of recent research in this field is that a valid high-level model must satisfy interventional consistency. This means that the act of simplifying a model and the act of intervening on a model must commute.
The Intuition: The Commutative Diagram
To understand why interventional consistency is necessary, imagine a system at two different levels of detail. The micro-level model contains fine-grained variables (e.g., individual cellular activity), while the macro-level model contains coarse-grained variables (e.g., organ-level function).
We can move between these worlds using two types of mappings:
- A state abstraction ($\tau$) that aggregates low-level states into high-level states.
- An intervention abstraction ($\omega$) that translates a high-level action into its corresponding low-level implementation.
Interventional consistency requires that if you start with a low-level state, there are two paths to predict the outcome of an intervention, and both paths must yield the exact same result:
Path 1: Intervene Micro-Level ──> Abstract State ──> Final Macro State
▲
│ (Must Match!)
│
Path 2: Abstract State ───────> Intervene Macro-Level ──┘
- Path 1 (The Ground Truth): You perform a detailed intervention directly on the micro-model, let the system evolve according to its complex rules, and then use $\tau$ to aggregate the final micro-state into a macro-state.
- Path 2 (The Abstract Shortcut): You abstract the initial micro-state to the macro-level using $\tau$, and then apply the simplified macro-intervention using $\omega$ directly within the macro-model.
As first formalized by Paul K. Rubenstein et al. in Causal Consistency of Structural Equation Models, a macro-level model is causally consistent if and only if these two paths commute. If they do not, the high-level model is a dangerous fiction: it might accurately describe passive observations, but it will fail to predict the consequences of real-world manipulations.
To prevent arbitrary or non-physical alignments between models, Sander Beckers et al. demonstrate in Abstracting Causal Models that a truly faithful abstraction requires the intervention mapping $\omega$ to be derived directly and strictly from the state mapping $\tau$.
Bridging Different Mathematical Worlds
Causal abstraction does not just apply to mapping one Structural Causal Model (SCM) to another; it can bridge entirely different mathematical formalisms.
- From Physics to SCMs: Real physical systems are often continuous-time processes governed by differential equations. Paul K. Rubenstein et al. proved in From Deterministic ODEs to Dynamic Structural Causal Models that interventions applied to underlying differential equations are mathematically equivalent to interventions applied to derived, high-level structural equations, providing a rigorous bridge from physics to causal graphs.
- From Actions to Programs: In computer science, we often want to reason about the causes of software behavior. Bita Banihashemi et al. extended abstraction theory in From Actions to Programs as Abstract Actual Causes, showing how low-level primitive actions can be abstracted into complex, non-deterministic programs while preserving the soundness and completeness of causal explanations.
- Handling Non-Causal Constraints: Real-world systems often contain non-causal relationships, such as mathematical definitions or unit conversions (e.g., converting Celsius to Fahrenheit). To accommodate this, Sander Beckers et al. introduced explicit constraints and "disconnect" operations in Causal Models with Constraints, expanding the types of systems that can be rigorously abstracted across scales.
Learning Abstractions from Data
In practice, we rarely know the exact structural equations of either the micro- or macro-models. A major thrust of recent research focuses on how to learn these consistent abstraction maps directly from data.
When multiple interventional distributions are available, Fabio Massimo Zennaro et al. demonstrated in Jointly Learning Consistent Causal Abstractions Over Multiple Interventional Distributions that differentiable programming can be used to optimize and discover the unknown mappings between variable distributions across scales. For simpler systems, Riccardo Massidda et al. established the precise graphical and parametric conditions required to guarantee valid abstractions in Learning Causal Abstractions of Linear Structural Causal Models.
However, gathering interventional data can be expensive or unethical. To circumvent this, researchers have developed methods to learn abstractions from purely observational data:
- Optimal Transport (COTA): In Causal Optimal Transport of Abstractions, Fabio Massimo Zennaro et al. introduced a framework that blends optimal transport with do-calculus constraints, allowing researchers to align micro- and macro-models without knowing the underlying equations.
- The Semantic Embedding Principle (SEP): In Causal Abstraction Learning based on the Semantic Embedding Principle, Gabriele D'Acunto et al. framed the discovery of abstraction maps as a geometric optimization problem on the Stiefel manifold, proving that consistent macro-level descriptions can emerge purely from the geometry of observational data.
Why It Works: Practical Applications
Enforcing interventional consistency is not just a theoretical exercise; it solves critical bottlenecks in simulation, decision-making, and machine learning.
1. Fast, Reliable Simulation Surrogates
Policy-makers often rely on Agent-Based Models (ABMs) to simulate complex systems like housing markets or disease spread. These simulations are highly detailed but computationally agonizing to run. Joel Dyer et al. showed in Interventionally Consistent Surrogates for Complex Simulation Models that by training a simplified surrogate model to maintain commutativity with the high-fidelity simulator, the surrogate can instantly and accurately predict the outcomes of novel policy interventions that were completely absent from its training data.
2. Accelerating Decision-Making
In complex multi-armed bandit problems, evaluating actions in a high-fidelity environment is costly. In Using causal abstractions to accelerate decision-making in complex bandit problems, Joel Dyer et al. developed the AT-UCB algorithm. By leveraging causal abstraction to prune sub-optimal actions in a cheap, coarse-grained model first, they proved that agents can drastically reduce cumulative regret when deployed in the expensive, high-fidelity environment.
3. Faithful Machine Learning Explanations
When explaining the decisions of black-box neural networks, standard statistical explanations often suffer from a lack of "faithfulness." Mette Friis Andersen et al. argue in Three Desiderata for Faithfulness in Machine Learning Explanations: The Case for Causal Abstraction that causal abstraction provides the exact mathematical framework needed to evaluate faithfulness, ensuring that an explanation captures true interventionist relations.
As a practical realization of this, Armin Kekić et al. introduced nonlinear Targeted Causal Reduction (nTCR) in Learning Nonlinear Causal Reductions to Explain Reinforcement Learning Policies. By applying small perturbations to a Reinforcement Learning agent's actions, nTCR extracts human-readable, macro-level causal models that accurately reflect the true underlying policy mechanics, exposing hidden biases that standard correlation-based explanations miss.
4. Dynamic Knowledge Representation
Finally, as autonomous agents encounter new data, they must update their understanding of the world without destroying what they have already learned. Matej Zecevic et al. introduced Continual Causal Abstractions (CCA) in Continual Causal Abstractions. CCA ensures that when an agent refines a model (e.g., zooming in from "total cholesterol" to specific "HDL/LDL" levels), the new, detailed logic maintains commutativity with its predecessor. This guarantees that updates to an agent's knowledge base never invalidate the functional utility of its earlier, simpler abstractions.
Key Foundational Work
Causal Consistency of Structural Equation Models (2017) — contributed 3 concept elements to this direction.
Counterfactual Causality from First Principles? (2017) — contributed 1 concept elements to this direction.
From Deterministic ODEs to Dynamic Structural Causal Models (2018) — contributed 1 concept elements to this direction.