🤖 AI

Intuitionistic $j$ -Do-Calculus in Topos Causal Models

This paper generalizes Pearl's do-calculus to an intuitionistic setting within Topos Causal Models by introducing $j$ -do-calculus, a sound rule system that defines causal interventions and conditional independences as local truths stable under a Lawvere-Tierney topology $j$ and interprets them via Kripke-Joyal semantics.

Original authors: Sridhar Mahadevan

Published 2026-07-28

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Sridhar Mahadevan

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 the clues you find depend entirely on which room of the house you are standing in. In one room, a fingerprint on the window proves the suspect entered; in the next room, that same fingerprint is just a smudge from a cleaning crew. This is the heart of causal inference, a branch of science that tries to figure out what causes what, rather than just what happens to happen together. Usually, detectives use a set of logical rules (called do-calculus) to ask "What would happen if I forced the suspect to enter?" by mentally cutting the usual chains of events. But these rules assume you are looking at the whole world from a single, perfect viewpoint.

The problem is that the real world isn't a single room; it's a building with many different contexts, or "regimes." A medical treatment might work in a sunny hospital but fail in a rainy clinic. A rule that works for one group of people might break for another. Scientists have long wanted a way to mix these different viewpoints together without losing the logic of the cause-and-effect chain. They need a system that can say, "This rule holds true in these specific rooms, and because these rooms cover the whole building, the rule holds true for the building too."

This is exactly what Sridhar Mahadevan's paper, Intuitionistic j-Do-Calculus in Topos Causal Models, sets out to do. The author builds a new mathematical "glue" that allows us to take the standard rules of cause-and-effect and apply them piece-by-piece across different environments, then stitch those pieces back together into a single, reliable truth.

The Detective's Toolkit: From Single Rooms to a Whole City

To understand the paper's magic, let's first look at the tools it uses. Imagine a city where every neighborhood has its own set of laws. In the "Observation District," you can only watch what happens naturally. In the "Intervention District," you can go around and physically move things to see what changes.

In the old way of doing things (called classical do-calculus), you had to pick one district and stick to it. If you wanted to know if a new policy caused a drop in crime, you had to assume your data came from a single, consistent world. But what if your data is a messy mix? What if some data comes from a time when it rained, and some from when it was sunny?

The paper introduces a new concept called j-stability. Think of this as a "safety net" for your logic. Instead of asking, "Is this rule true everywhere?" the paper asks, "Is this rule true on a specific set of local maps that, when put together, cover the whole territory?"

Here is how the paper's new system works, using a playful metaphor:

1. The Map of Regimes (The Site)
Imagine the world is a giant puzzle, but instead of flat pieces, the pieces are different "regimes" or contexts. One piece is "Observational Data," another is "Experimental Data," and another is "Data from a Specific Hospital." In the paper's language, these are called stages or charts. The author treats these not just as separate piles of data, but as a connected network where you can move from one context to another.

2. The Local Rules (Chartwise Semantics)
On each individual puzzle piece (each regime), the author uses the standard, trusted rules of cause-and-effect. If you are in the "Observational" piece, you use the standard logic to see if two things are independent. If you are in the "Intervention" piece, you use the standard logic to see what happens when you force a change. The paper insists that we don't invent new rules for these local pieces; we just use the ones we already trust.

3. The Glue (Sheaf Theory and Descent)
This is the most creative part. The paper asks: "If a rule holds true on the 'Observational' piece AND it holds true on the 'Interventional' piece, and these two pieces overlap in a way that makes sense, can we say the rule is true for the whole puzzle?"

The answer is yes, but only if the pieces fit together perfectly. The author uses a mathematical concept called sheaf theory (think of it as a very strict, high-quality glue) to ensure that the local truths don't contradict each other. If the local truths agree on the overlaps, the "glue" allows them to descend into a single, global truth. This is called descent.

What the Paper Actually Finds

The paper proves a local-to-global soundness theorem. In plain English, this means:

If you can verify a causal rule (like "Variable A causes Variable B") on a collection of local maps (a J-cover) that together cover your entire situation...
And if the way you move between these maps preserves the causal structure (like keeping the same variables and rules)...
Then you are mathematically guaranteed that the rule holds true for the entire situation, even if you can't see the whole picture at once.

The author applies this to Pearl's three famous rules of do-calculus. These rules tell us when we can swap an action (like a drug) for an observation (like a symptom) in our equations. The paper shows that these rules still work in this messy, multi-regime world, provided they hold true on the local maps.

What the Paper Explicitly Rejects

It is crucial to know what this paper does not do, because the author is very careful to draw boundaries:

It does not claim to solve everything. The paper explicitly states it is a "soundness" theorem, meaning it tells you when you are safe to use the rules. It does not claim to be a "completeness" theorem, which would mean it can find every possible causal answer. There might be causal truths that this method simply cannot reach.
It does not say "Intervention" and "Conditioning" are the same thing. In some advanced math, people try to treat "doing" something and "watching" something as two sides of the same coin (using a concept called Kan extensions). This paper explicitly rejects that idea. It argues that intervention is a physical replacement of a mechanism (like swapping a part in a machine), while conditioning is just looking at a specific slice of data. They are different operations, and the paper refuses to blur them together.
It does not assume the world is always "Boolean" (black and white). In standard logic, a statement is either true or false. In this new framework, a statement can be "true on a cover" (true in the local maps) without being "globally true" in a simple, binary sense. The logic is more flexible, like a dimmer switch rather than an on/off switch.

The Bottom Line: A New Way to Stitch the World Together

The paper doesn't give you a new drug or a new algorithm to run on your computer right now. Instead, it gives you a new logical framework for how to think about causal reasoning when the world is complex and fragmented.

Imagine you are trying to predict the weather. You have a local forecast for the beach, one for the mountains, and one for the city. If the beach forecast says "sunny" and the mountain forecast says "sunny," and they overlap in a way that makes sense, you can confidently say "It's sunny everywhere." But if the beach says "sunny" and the mountain says "rainy," you can't just guess.

Mahadevan's paper provides the rigorous mathematical proof for when you are allowed to stitch those local forecasts together into a global prediction. It proves that if the local rules of cause-and-effect hold up in every specific context (the "charts"), and those contexts cover the whole picture, then the causal conclusion is valid for the whole system.

The author is careful to note that this is a theoretical breakthrough. The paper assumes you already have the local maps and the covers; it doesn't tell you how to find them from raw data (that's a job for a companion paper). But for anyone trying to build AI or machine learning systems that need to reason about cause and effect across different environments—like a self-driving car that behaves differently in rain versus sunshine—this paper provides the essential "glue" to make sure the logic holds together.

In short, the paper says: "You don't need a single, perfect view of the world to know the truth. If you have enough local views that fit together, the truth will emerge from the glue."

Technical Summary: Intuitionistic j-Do-Calculus in Topos Causal Models

Problem Statement
Pearl's do-calculus provides a complete axiom system for interventional identification in acyclic structural causal models (SCMs) under classical Boolean logic. However, standard do-calculus assumes a static, global context. In many real-world scenarios, causal structures and data availability vary across different regimes, experimental contexts, or "stages" (e.g., different laboratories, policy environments, or time periods). The paper addresses the challenge of formalizing causal inference in such regime-indexed settings where causal assertions may hold locally but not globally, and where the logic governing these assertions may be intuitionistic rather than classical. The core problem is to extend do-calculus to a setting where causal models vary over a site (a category of regimes) and where the validity of causal rules depends on the existence of a covering family of local charts.

Methodology
The authors develop a sheaf-valued, regime-indexed semantics for Pearl's do-calculus within the framework of Topos Causal Models (TCMs). The methodology proceeds through several categorical and logical layers:

Categorical Framework: Causal models are treated as objects in a presheaf topos $\mathbf{Set}^{\mathcal{C}^{op}}$ over a small category $\mathcal{C}$ (the site of regimes). A Grothendieck topology $J$ on $\mathcal{C}$ specifies which families of morphisms (charts) constitute a "cover," representing admissible local views of a regime.
Sheaf Semantics: The authors work within the topos of sheaves $\mathbf{Sh}_J(\mathcal{C})$ . A causal assertion is defined as $J$ -stable (or $j$ -stable) at a stage $U$ if it holds on a $J$ -cover of $U$ in the sense of Kripke–Joyal semantics. This means a statement is "true" at $U$ if it can be verified on a family of local charts that jointly cover $U$ and are compatible under restriction.
Intervention and Conditioning:
- Interventions are modeled as structural-mechanism or stochastic kernel replacements (surgical operations on the causal graph), formalized via the Kleisli category of a distribution monad ( $\mathbf{Kl}(\mathbf{Dist})$ ).
- Conditioning is defined strictly when the required finite-discrete conditional has positive normalizing mass (or via a compatible disintegration), implemented as restriction to a comprehension subobject followed by normalization.
- The paper explicitly rejects identifying intervention and conditioning as a universal pair of left and right Kan extensions, instead emphasizing sheafification as reflective localization and descent.
Logic: The internal logic of the topos is intuitionistic (Mitchell-Bénabou language). The validity of causal rules is determined by $J$ -forcing: a rule applies if its premises (e.g., $d$ -separation conditions) hold on a $J$ -cover.

Key Contributions
The paper makes four primary theoretical contributions:

Definition of $J$ -Local $d$ -Separation and Interventional Equality: The authors define $j$ -d-separation ( $X \perp^j_U Y | Z$ ) as the condition where a conditional independence holds on a $J$ -cover of a stage $U$ . This lifts the classical notion of $d$ -separation to a local-to-global framework.
Compatibility Assumptions for Sheaf Descent: The paper states explicit assumptions under which chartwise SCM semantics forms a sheaf of causal models. Specifically, restriction maps must preserve variable signatures, graph surgeries (mutilations), causal kernels, and the conditionals used by the calculus.
Local-to-Global Soundness Theorem: The central result is a soundness theorem for regime-aware causal reasoning. It proves that if Pearl's three rules are applied chartwise on a $J$ -cover, and the resulting interventional kernels agree on overlaps, then the equality descends to the covered stage $U$ via the sheaf separatedness axiom. This yields $J$ -local forms of Pearl's three rules ( $J_1, J_2, J_3$ ) without altering their original mutilated-graph premises.
Conservativity and Reductions: The framework demonstrates two key reductions:
- The terminal site ( $C=1$ ) recovers ordinary set-based do-calculus.
- The trivial Grothendieck topology recovers presheaf semantics (stagewise logic), though the internal logic of a general presheaf topos need not be Boolean.

Results and Derivations
The paper provides worked derivations illustrating how the three rules of do-calculus translate into internal equalities within a TCM:

Rule 1 (Insert/Delete Observations): If $Y \perp Z | X, W$ in the mutilated graph $G_{\bar{X}}$ on a $J$ -cover, then $P(Y | do(X), Z, W) = P(Y | do(X), W)$ holds internally.
Rule 2 (Action/Observation Exchange): If $Y \perp Z | X, W$ in $G_{\bar{X}, \underline{Z}}$ on a $J$ -cover, then intervening on $Z$ is equivalent to observing $Z$ .
Rule 3 (Insert/Delete Actions): If $Y \perp Z | X, W$ in $G_{\bar{X}, \underline{Z(W)}}$ on a $J$ -cover, then $do(Z)$ is irrelevant.

The authors illustrate these concepts using simple DAGs (Earthquake/Burglary and Pollution/Smoker/Cancer), showing how a conditional independence that fails globally (e.g., conditioning on a collider) can become $J$ -stable if the cover includes an interventional chart that cuts the collider's incoming edges.

Significance and Claims
The paper positions itself as a conceptual work that establishes the logical foundations for causal inference in varying regimes.

Scope: It assumes the theoretical objects (stages, covers, internal distributions) are given. It does not provide algorithms for data-driven construction of these objects, noting that this is addressed in companion work.
Soundness vs. Completeness: The authors explicitly state they make no completeness claim for the resulting regime-indexed calculus. The contribution is a soundness theorem: valid local reasoning glues to valid global reasoning under specific compatibility conditions.
Categorical Distinctions: The work clarifies the categorical role of sheafification (descent) versus Kan extensions. It argues that while conditional expectation has a right Kan extension interpretation, causal intervention is fundamentally a mechanism replacement, and no universal left/right Kan duality exists between them without further indexed structure.
Extension to Exchangeability: The paper briefly extends the framework to exchangeable j-stable causality, showing how $j$ -stability interacts with permutation invariance, recovering "do-Finetti" principles internally when the topology is invariant under group actions.

In summary, the paper provides a rigorous, sheaf-theoretic generalization of Pearl's do-calculus, enabling causal reasoning that is stable across regimes and compatible with intuitionistic logic, while strictly delineating the boundaries between local verification and global truth.

The Detective's Toolkit: From Single Rooms to a Whole City

What the Paper Actually Finds

What the Paper Explicitly Rejects

The Bottom Line: A New Way to Stitch the World Together

More like this