🤖 machine learning

Beyond Directed Acyclic Graphs: Causal Zeros and Causal Differential Equations

This paper extends Pearl's structural causal model framework beyond directed acyclic graphs by formalizing symmetric constraints as "causal zeros" and grounding feedback loops in Causal Differential Equations, thereby enabling causal reasoning for systems with non-directional relationships and dynamic cycles.

Original authors: Sergei V. Kalinin

Published 2026-07-28

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Sergei V. Kalinin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but instead of looking for clues, you are trying to figure out what caused what. In the world of science and data, this is called causal inference. For a long time, the best tool detectives had was a set of rules built on "Directed Acyclic Graphs" (DAGs). Think of a DAG like a one-way street map. It tells you that if you push a domino (the cause), it will knock over the next one (the effect), but it strictly forbids the domino from knocking the first one back over. This works perfectly for things like "smoking causes lung cancer" or "rain makes the grass wet." The direction is clear, and the map is a straight line with no loops.

However, the real world is messy. Sometimes, things happen in a circle, or two things just exist together in a perfect balance without one clearly "pushing" the other. Think of a seesaw: if you go up, the other person goes down. But did your weight cause them to go down, or did their weight cause you to go up? In physics and economics, many laws work like this—they are symmetric constraints where the direction only appears when you decide to push on one side. The big question this paper tackles is: How do we build a detective's map for these circular, balanced, or "two-way" relationships without breaking the rules of logic?

The Problem with One-Way Streets

The paper starts by pointing out a crack in the foundation of modern causal reasoning. The current "gold standard" framework, developed by Judea Pearl, is fantastic for straight lines, but it hits a wall when it meets symmetric constraints.

Imagine the Ideal Gas Law, the famous equation $PV = nRT$. It relates Pressure ( $P$ ), Volume ( $V$ ), and Temperature ( $T$ ). In a physics lab, you can squeeze a gas (change $V$ ) and watch the pressure ( $P$ ) rise. Or, you can heat it up (change $T$ ) and watch the pressure rise. The equation works both ways. But the current "one-way street" maps can't handle this. They demand you pick a direction before you start. They force you to say, "Pressure causes Volume," or "Volume causes Pressure." But in reality, neither is the boss; they are just partners in a dance. If you force a direction where there isn't one, you might make a prediction that breaks the laws of physics.

The authors argue that for many physical systems—like gas laws, supply and demand in economics, or electrical circuits—the relationship isn't a one-way arrow. It's a symmetric constraint. It's a rule that says, "These variables must fit together like puzzle pieces," but it doesn't say which piece is the driver and which is the passenger.

Enter the "Causal Zero"

To fix this, the authors introduce a new character to the story: the Causal Zero.

Think of a Causal Zero not as a street, but as a magnetic field or a tensioned rubber sheet. It's a relationship that holds things together but has no direction of its own. In their new "Extended Causal Model," these relationships are drawn as hyperedges (a fancy term for a multi-way connection) instead of arrows.

Here is the magic trick: A Causal Zero sits there, neutral and directionless, until you perform an intervention. An intervention is like a scientist stepping in and saying, "I am going to hold this variable fixed."

The paper proposes a new tool called the Activation Operator. This is like a switch that flips the neutral rubber sheet into a one-way street only when you need it to.

The Setup: You have a Causal Zero connecting Pressure, Volume, and Temperature.
The Intervention: You decide to clamp the Volume (you fix it so it can't change).
The Activation: The system asks, "Okay, since Volume is fixed, which variable are we solving for?" If you say, "Solve for Pressure," the rubber sheet snaps into a one-way arrow: Volume $\to$ Pressure. If you say, "Solve for Temperature," the arrow flips: Volume $\to$ Temperature.

The crucial insight here is that the direction doesn't exist until you specify what you are fixing (the driver) and what you are solving for (the target). Without naming the target, the system is stuck in limbo. The authors emphasize that you can't just say "I'm intervening on Volume"; you must also say "and I want to know what happens to Pressure." If you forget the target, the math breaks.

Unrolling the Time Loop

The paper also tackles the second big problem: Cycles.
Standard maps hate loops. They say, "No circles allowed!" But in the real world, feedback loops are everywhere. A thermostat measures the temperature, turns on the heater, which warms the room, which the thermostat measures again. It's a circle.

The authors suggest a clever way to handle this for physical systems: Time.
They argue that what looks like a magic instant loop is actually just a fast-moving process that we are watching too slowly. If you zoom in with a super-fast camera, you see that the temperature change causes the heater to turn on a tiny fraction of a second later. The loop isn't a circle; it's a spiral that moves forward in time.

To explain this, they introduce Causal Differential Equations (CDEs).

The Transient Phase: When things are changing (like the gas heating up), the system is a straight line of cause-and-effect moving through time. It's a "time-unrolled" acyclic process.
The Equilibrium Phase: Once the system settles down (the gas stops changing), the time stops mattering, and you are left with the static Causal Zero (the equation $PV=nRT$).

So, the paper suggests that the "loop" is just an artifact of ignoring time. If you include time, the loop unravels into a long, straight line of cause-and-effect. The Causal Zero is just the snapshot of that line when it finally comes to rest.

What This Means for the Future

The authors are careful to say this is a proposal and a framework, not a finished, solved puzzle. They have built the rules for how to draw these new maps and how to "activate" them, but they admit there are still open questions.

They suggest that for complex, chaotic systems (like weather or a bouncing ball in a chaotic box), the "equilibrium" might not be a single point but a wild, looping attractor. They propose that we might need to define interventions relative to these loops, but that is still an open problem.
They note that while their "Activation Operator" works for many physical laws, figuring out how to discover these Causal Zeros from raw data (without knowing the physics beforehand) is a challenge for future research.

The Takeaway

In simple terms, this paper says: "Stop trying to force every relationship into a one-way arrow. Some relationships are just rules that hold things together."

They give us a new way to draw these rules: as neutral "Causal Zeros" that only become one-way arrows when we explicitly decide which variable we are controlling and which one we are measuring. And for the loops that confuse us, they remind us to look at the clock: if you watch fast enough, the circle is just a line moving forward.

It's a way to make the math of cause-and-effect fit the messy, balanced, and sometimes circular reality of the physical world, without breaking the logic that makes the math work in the first place.

Technical Summary: Beyond Directed Acyclic Graphs: Causal Zeros and Causal Differential Equations

Problem Statement
Judea Pearl's Structural Causal Model (SCM) framework, grounded in Directed Acyclic Graphs (DAGs) and the do-calculus, serves as the dominant formal language for causal reasoning. However, the paper identifies two structural limitations that prevent this framework from adequately modeling certain physical and economic phenomena:

Non-Causal Symmetric Relationships: Many fundamental physical laws (e.g., the ideal gas law $PV=nRT$, conservation laws, equilibrium constraints) are intrinsically symmetric. They do not possess an intrinsic causal direction; rather, the direction of causality emerges only upon intervention. In standard SCMs, the causal graph must be pre-specified with directed edges, making it impossible to represent these symmetric constraints without arbitrarily assigning a direction that may not exist in the underlying physics.
Feedback Cycles: The SCM framework strictly forbids cycles. While feedback is ubiquitous in dynamical systems (e.g., gene regulation, thermostats), standard discovery algorithms and do-calculus rules rely on acyclicity. Existing approaches to cyclic models often treat them as static objects or require complex solvability conditions, potentially obscuring the temporal nature of the feedback.

Methodology
The author proposes an Extended Causal Model (ECM) that integrates two primary extensions to the standard Pearl framework:

Causal Zeros: The paper introduces the "causal zero" as a first-class object. A causal zero is a symmetric, lawlike constraint (a hyperedge) among variables that holds at equilibrium but carries no intrinsic causal orientation.
- Activation Operator: To resolve a causal zero into a directed mechanism, an "activation operator" is defined. This operator requires specifying both a driver (the variable externally fixed) and a target (the variable solved for).
- Compilation: Upon activation, the symmetric constraint is compiled into a standard directed structural equation ( $X_{target} := g(X_{driver}, \dots)$ ), effectively converting the hyperedge into a directed edge within a DAG. This process is subject to local solvability (implicit function theorem) and graph-admissibility (ensuring no new cycles are formed) conditions.
Causal Differential Equations (CDEs): To ground these static constraints and cycles in physical reality, the paper introduces Causal Differential Equations.
- Time Unrolling: Apparent instantaneous cycles in static models are treated as artifacts of suppressed time. In CDEs, feedback is represented as temporal coupling, where the system evolves through a transient, acyclic, time-unrolled causal process.
- Equilibrium Manifolds: Causal zeros are formally identified as the defining functions of attracting equilibrium manifolds of the underlying dynamical system. When a system relaxes to a stationary state, the dynamic equations reduce to the symmetric constraint (the causal zero).
- Intervention Semantics: The framework distinguishes between hard interventions (equation replacement), soft interventions (high-gain forcing), and transient interventions, showing how these map to the dynamics of the CDE.

Key Contributions

Formalization of Causal Zeros: The paper defines causal zeros as tuples comprising constraint variables, a smooth scalar constraint function, and sets of admissible drivers and targets. It establishes that a single scalar constraint can determine exactly one variable given the others, necessitating the explicit naming of a target variable during intervention.
The Activation Operator: A novel operator is defined that compiles an undirected constraint into a directed structural equation. The paper emphasizes that omitting the explicit target label renders the operation ill-posed for constraints involving more than two variables, correcting a potential ambiguity in prior proposals.
Unified Dynamic Framework: The paper provides a rigorous correspondence (Theorem 6.1) between Causal Differential Equation Systems (CDES) and causal zeros. It demonstrates that regular SCMs are a special case where dynamics relax to an isolated stable point, while causal zeros arise when the equilibrium set has positive dimension (a manifold).
Extended Do-Calculus: The paper outlines "Rule 0" for activation, which allows standard do-calculus rules (Rules 1–3) to be applied post-activation. It also discusses counterfactual semantics within this extended framework, distinguishing between contrasts within a single activated law and cross-model comparisons.

Results and Theoretical Findings

Symmetry vs. Direction: The paper demonstrates that symmetric physical constraints (Type I: equilibrium laws, Type II: conservation laws) are not merely "missing" arrows but are fundamentally different objects that require intervention to acquire direction.
Cycle Resolution: For finite-propagation, well-posed state-space systems, cycles are shown to be artifacts of time suppression. Restoring time via CDEs yields an acyclic process where the "cycle" is a projection of temporal feedback.
Relation to Prior Work: The proposed framework refines, rather than replaces, Causal Constraints Models (CCMs) and labeled equilibrium equations. Unlike CCMs, which treat activation sets as admissibility filters, the proposed activation operator explicitly compiles the constraint into a predictive structural equation.
Limitations and Open Problems: The paper acknowledges that the completeness of the extended do-calculus (Rules 0–3) for ECMs remains an open question. It also notes that for systems with limit cycles or strange attractors, a theory of "attractor-relative interventions" (phase-dependent responses) is required but not yet fully developed.

Significance
The paper claims significance by addressing a conceptual gap in causal inference where standard formalisms fail to accommodate symmetric physical laws and feedback loops without arbitrary assumptions. By introducing causal zeros and grounding them in causal differential equations, the framework:

Provides a mathematically consistent way to model equilibrium physics and economic identities without pre-committing to a causal direction.
Bridges the divide between static causal graphs and dynamical systems, showing how the former emerges from the latter under specific relaxation conditions.
Offers a more faithful representation of scientific modeling in fields like materials science and thermodynamics, where "intervention" often dictates the causal direction of a symmetric constraint.

The author positions this work as a modest refinement of existing causal programs (specifically CCMs and labeled equilibrium SCMs), offering a specific "orientation operator" that correctly handles target labeling and admissibility, thereby enabling the application of standard causal calculus to a broader class of physical and economic systems.