Causal Reasoning in Stochastic and Nondeterministic Systems

41 concepts36 papers3 proposals

Traditional causal inference relies heavily on a foundational assumption: determinism. In standard Structural Causal Models (SCMs), every variable is modeled as a strict, deterministic function of its direct causes and some unobserved background noise. Under this view, uncertainty is not inherent to the universe; it is merely a symptom of our ignorance about hidden variables. If we could perfectly observe the background state, every effect would be completely predictable.

However, this deterministic assumption creates significant friction when applied to modern, real-world systems. Consider a large language model generating text, a cyber-physical system operating over a lossy network, or a complex public policy initiative. In these settings, treating relationships as perfectly deterministic functions is either computationally intractable or conceptually flawed. We cannot easily access or control the internal random seeds or microscopic background variables of these systems.

To bridge this gap, a major direction of research focuses on extending causal modeling to inherently stochastic and nondeterministic environments. This shift requires rewriting the rules for how we represent systems, how we intervene on them, and how we assign actual blame or responsibility for specific outcomes.

Redefining Actual Causation Under Indeterminism

A core challenge in causal reasoning is determining actual causation (or token causality)—the task of proving whether a specific event caused a specific outcome in a singular instance (e.g., "Did this patient's smoking cause their lung cancer?").

In a deterministic world, actual causation is often evaluated using "but-for" counterfactual logic: if the cause had not occurred, the effect would not have happened. But as Christopher A. Merck et al. demonstrate, this binary logic breaks down in stochastic systems. If an effect can occur through alternative random paths by sheer chance, the effect might still happen even in the absence of the cause.

To resolve this, researchers have moved toward frameworks like Causal Probabilistic logic (CP-logic) and Nondeterministic Structural Causal Models (NSCMs). As shown by Sander Beckers et al., translating historically deterministic causal formalisms into probabilistic logic allows models to seamlessly handle non-deterministic scenarios and causation by omission.

Deterministic Counterfactual:
[Cause: Yes] ---> [Effect: Yes]
[Cause: No ] ---> [Effect: No ]  (Clear "But-For" Dependency)

Stochastic Counterfactual:
[Cause: Yes] ---> [Effect: 80% Chance]
[Cause: No ] ---> [Effect: 20% Chance]  (Dependency is a Shift in Probability)

Instead of requiring access to hidden deterministic variables or exact random seeds, Sander Beckers demonstrates that we can evaluate actual causation by looking for counterfactual dependence within a "structurally simplified" version of the world. This insight allows us to treat highly complex systems—such as Large Language Models—as inherently nondeterministic causal models, enabling valid counterfactual reasoning simply by re-running the model with alternative prompts rather than tracking internal weights or random seeds (Sander Beckers).

When outcomes are unpredictable, assigning responsibility becomes a matter of tracking execution paths and probability shifts. Shakil M. Khan 0001 et al. extended these logic frameworks to nondeterministic domains where agents lack full control over their actions, providing a formal mathematical basis for determining who or what is to blame when a system fails. This is crucial for diagnosing failures in stochastic environments like data centers (Yang Wu) or autonomous systems operating under uncertainty (Sander Beckers et al.).

From Atomic Overrides to Stochastic Interventions

In Pearl’s traditional causal calculus, an intervention is modeled using the $do$-operator (e.g., $do(X = x)$). This represents an "atomic" intervention: a hard, deterministic override that completely severs a variable from its natural causes and forces it to a fixed constant value.

While mathematically clean, atomic interventions fail to capture the nuance of real-world policies. For example, a medical guideline does not force every patient to take an exact dosage; instead, it alters the probability distribution of treatments based on individual patient characteristics.

To formalize these scenarios, Juan D. Correa et al. introduced the $\sigma$-calculus, a powerful generalization of the $do$-calculus designed specifically for stochastic and conditional interventions.

Instead of replacing a variable's structural equation with a fixed constant, a stochastic intervention replaces it with a new conditional probability distribution. The $\sigma$-calculus provides a complete set of graphical rules to determine whether the effects of such probabilistic policies can be identified from existing observational or experimental data.

Atomic Intervention:      do(X = x)   ---> Forces X to a single constant
Stochastic Intervention:  σ(P(X|Z))   ---> Modifies the probability distribution of X given Z

This mathematical framework allows decision-makers to evaluate complex, realistic "soft" policies. For instance, Alexis Bellot leverages this perspective to estimate tight bounds on policy effects even when unobserved confounders are present, treating interventions as continuous or conditional functions rather than binary switches. This ensures that causal characterizations remain robust to inherent randomness and latent variables without relying on functional determinism (Jin Tian et al.).

Modeling Causality via Probabilistic Programs

If the world is inherently stochastic, how should we represent the underlying structural equations? Traditional SCMs are structurally rigid, usually requiring a fixed, finite set of variables arranged in a directed acyclic graph.

To capture the open-ended, looping, or recursive nature of stochastic processes, researchers have turned to computer simulations and probabilistic programming. Duligur Ibeling formalized this by modeling causal systems as probabilistic Turing Machines that consume an infinite sequence of independent random bits. This approach allows for the representation of complex, recursive logic that exceeds the expressive power of traditional structural equations, providing a rigorous foundation for open-universe causal reasoning where the number of variables is not fixed in advance (Duligur Ibeling et al.).

In a similar vein, Joseph Halpern et al. introduced Generalized Structural-Equations Models (GSEMs). Instead of focusing on specific deterministic equations, GSEMs abstract the system into a high-level "interface" of interventions and outcomes, integrating probability distributions over contexts. This mathematical generalization allows standard causal concepts—like responsibility and blame—to be applied directly to complex dynamical systems, such as ordinary differential equations (ODEs) and hybrid automata, where the same intervention can lead to a range of probabilistic outcomes.

Key Foundational Work

Directed Constraint Networks: A Relational Framework for Causal Modeling (1991) — contributed 1 concept elements to this direction.
Formalizing Narratives Using Nested Circumscription (1996) — 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_stochastic_causal_modeling_25aea4c7,
  title = {stochastic causal modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/stochastic-causal-modeling/rdr_25aea4c7e25944ecbdca3a27d263b16f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_stochastic_causal_modeling_25aea4c7_v1,
  title = {stochastic causal modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/stochastic-causal-modeling/rdr_25aea4c7e25944ecbdca3a27d263b16f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake