📈 economics

Interventional Score Geometry for Causal Inference

This paper proposes a geometric framework for causal inference that extends observational score geometry to interventional settings by defining causal influence through the variation of marginal interventional score fields, thereby establishing a geometric dictionary for randomized trials and instrumental variables while clarifying the limitations of score-based identification within Pearl's Ladder of Causation.

Original authors: Mojtaba Eslami

Published 2026-07-27

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Mojtaba Eslami

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

The Map vs. The Movie

Imagine you are a detective trying to figure out who pushed whom in a crowded room. You have a high-resolution photo of the scene. In the photo, Person A is standing right next to Person B, and both look like they are leaning forward. Based on this single snapshot, you can say, "A and B are close together," but you cannot tell who pushed whom. Maybe A pushed B, maybe B pushed A, or maybe a third person pushed both of them at the same time. In the world of data science, this snapshot is called an observational distribution. It's just a picture of what things tend to happen together.

For decades, scientists have tried to use fancy math to turn that static photo into a movie, hoping to see the action unfold and figure out the cause-and-effect. They built beautiful mathematical maps called information geometry, which treat data like a landscape with hills and valleys. The idea was that if you studied the shape of the landscape carefully enough, the "push" (the cause) would reveal itself. But there's a catch: a photo of a river doesn't tell you which way the water is flowing; it just shows you where the water is. To know the direction, you need to see what happens when you throw a rock in. This paper is about realizing that you can't figure out the flow just by staring at the water; you have to actually throw the rock.

The Paper's Big Idea: Why the Photo Isn't Enough

This paper, titled Interventional Score Geometry for Causal Inference, argues that looking at data alone is like trying to understand a musical instrument by only listening to the sound it makes when no one is touching it. The authors, led by Mojtaba Eslami, show that the "shape" of the data (the observational geometry) is exactly the same whether the music was caused by a violin or a cello. If two different stories (causal models) create the exact same pattern of data, they will look identical on any map built just from that data.

The paper proves that you cannot magically extract the "cause" from the "effect" just by doing more complex math on the same old numbers. It's like trying to figure out if a river flows north or south by only looking at a still photo of the water; the photo is the same either way. To know the direction, you have to intervene. You have to reach in and change something.

The "Hard" Intervention: Clamping a String

The authors introduce a new way to think about changing the world, which they call a hard intervention. Imagine a guitar with six strings. If you pluck the strings, they vibrate and make a chord. That's the observational data. Now, imagine you take a clamp and hold the "A" string completely still at a specific note. You haven't just changed the volume of that string; you've removed it from the game entirely. The string can no longer vibrate, and the other strings might change how they sound because they are no longer interacting with it.

In the paper's math, this "clamping" is called do( $X_k = \xi$ ). It means you force a variable (like a drug dosage or a temperature) to be a specific number, $\xi$ , regardless of what it wanted to be before. The paper shows that when you do this, the math changes completely. You can no longer use the old map because the landscape has been flattened. The "score" (a mathematical arrow that points to where the data is most likely to be) now lives on a lower-dimensional space, like a 2D drawing of a 3D object.

The New Tool: The "Score" of the Change

The paper builds a new dictionary for understanding these changes. Instead of just looking at the static photo, they look at how the "score" (the direction of the data) changes as you twist the intervention knob.

They define causal influence simply: If you change the setting of $X$ and the behavior of $Y$ changes, then $X$ causes $Y$ . But they go deeper. They show that you can detect this cause by looking at a specific derivative (a rate of change) of the new, clamped data. If this "interventional score" is zero, there is no causal link. If it's not zero, there is a link.

To prove this works, they use a simple example with two variables, $X$ and $Y$ , that look like a standard bell curve (a Gaussian distribution). They create two different stories:

Story A: $X$ causes $Y$ .
Story B: $Y$ causes $X$ .

Both stories produce the exact same photo (the same data distribution). If you only look at the photo, you can't tell them apart. But when the authors apply their new "clamping" math, the results are totally different. In Story A, clamping $X$ changes the score of $Y$ . In Story B, clamping $X$ does absolutely nothing to $Y$ 's score. The math successfully separates the two stories, but only because they forced a change. The photo alone never could have told them apart.

What This Rules Out

The paper is very clear about what it doesn't do. It explicitly rules out the idea that you can just take a list of "allowed" changes (like "we can change the price of a product") and project the old data onto that list to find the answer. They prove that even if you know exactly which levers you can pull, the old data doesn't contain the answer to what happens when you pull them. You need extra information—structural knowledge or experimental data—to fill that gap.

They also warn against a common confusion in the field of machine learning. There are other methods called "score-based" models (like those used to generate fake images) that try to learn the shape of data. The paper says these models are great at learning the photo, but they cannot learn the movie (the causal structure) just by getting better at the photo. No matter how perfect the image generator is, it still can't tell you what happens if you push the button, unless you actually give it data from when the button was pushed.

The "Causal Metric": Measuring the Push

Finally, the authors create a new way to measure how sensitive a system is to a push. They call this a causal metric. Think of it like measuring how much a spring stretches when you pull it. If you pull it a tiny bit and it stretches a lot, the spring is very sensitive. If it barely moves, it's stiff.

The paper shows that you can measure this sensitivity using a standard statistical tool called Fisher information, but you have to be careful to measure it on the right "slice" of the data (the part that wasn't clamped). They show that for simple cases, like sliding a distribution along a line, this metric is just a constant number that tells you how "sharp" the data is. This helps researchers know how much data they need to collect to see a real effect. If the metric is small, you need a huge amount of data to see the change. If it's large, a small experiment will show the result clearly.

The Bottom Line

This paper doesn't give us a magic wand to find causes in any dataset. Instead, it gives us a precise vocabulary and a set of rules for understanding what we can and cannot know. It confirms that causality lives in the "what if" scenarios (the interventions), not in the "what is" scenarios (the observations).

The authors are careful to say they haven't solved everything. They haven't figured out how to predict what would have happened to a specific person if they had taken a different path (the "counterfactual" rung of the ladder). They've only built a solid bridge between the "what is" and the "what happens if we change it." They've shown that to understand the flow of the river, you have to stop staring at the water and start throwing rocks.

Technical Summary: Interventional Score Geometry for Causal Inference

Problem Statement
The paper addresses a fundamental limitation in causal inference: the inability to identify causal direction or causal effects solely from the joint observational density $p(x)$ of observed variables $X$ . The author argues that the "observational geometry"—constructed from the support $M$ , a background metric $g$ , and the score field $\psi(x) = \nabla_x \log p(x)$ —is entirely determined by $p(x)$ . Consequently, any two distinct structural causal models (SCMs) that generate the same reduced-form distribution $p(x)$ induce identical observational geometries. This mirrors the classical non-identification problem where distinct causal structures (e.g., $X \to Y$ vs. $Y \to X$ ) can be observationally equivalent. The paper posits that causal asymmetry cannot be derived from $p(x)$ alone but must be supplied via an explicit intervention.

Methodology and Framework
The paper develops a geometric framework that distinguishes between observational and interventional states, aligning with Pearl's "Ladder of Causation."

Interventional Score Definition:
The core technical innovation is the definition of an interventional score. A hard intervention $do(X_k = \xi)$ fixes a coordinate $X_k$ to a value $\xi$ , collapsing the joint density onto a $(d-1)$ -dimensional submanifold $\{x_k = \xi\}$ . Because the resulting distribution has no density with respect to the original $d$ -dimensional volume, a naive gradient is ill-defined.
- Definition 1: The interventional score $\psi^{(k, \xi)}_{-k}$ is defined strictly on the remaining non-intervened coordinates $X_{-k}$ . It is the gradient of the log-density of the marginal interventional distribution $p^{(k, \xi)}_{-k}(x_{-k})$ .
Causal Influence and Diagnostics:
- Definition 2 (Distributional): Causal influence $X_k \leadsto X_j$ is defined as the variation of the marginal law of $X_j$ under $do(X_k = \xi)$ as $\xi$ changes.
- Definition 3 (Score-Detectable Influence): A sufficient condition for causal influence is the non-vanishing of the mixed derivative of the log-marginal density: $S_{k \to j}(\xi, x_j) = \frac{\partial}{\partial \xi} \frac{\partial}{\partial x_j} \log p^{(k, \xi)}_j(x_j) \neq 0$ .
- Proposition 2: On a fixed, connected, strictly positive support, this score diagnostic is also necessary. If the score derivative is zero everywhere, the marginal distribution does not change with the intervention.
The Admissible Set and Response Field:
The paper rejects the notion that an "admissible set" of intervention directions $A(x)$ (the set of manipulable coordinates) is sufficient to recover causal effects from observational data.
- Proposition 3: Two models can share the same observational score $\psi$ and the same admissible set $A(x)$ yet possess different causal responses. Therefore, projecting $\psi$ onto $A(x)$ cannot yield a causal field.
- Definition 4 (Interventional Response Field): Causal content is located in a response field $S(v)$ , which describes the sensitivity of the interventional score to a specific structural intervention family. This field must be supplied by external structural knowledge or experimental design; it cannot be inferred from $p(x)$ and $A(x)$ alone.
Causal Metric:
To measure the "distance" or information content of interventions, the paper defines a Causal Metric (Definition 5). Unlike naive constructions that attempt to compare sensitivities across different targets (which live on different coordinate spaces), this metric is built as the Fisher information metric on a family of interventions sharing a common target $k$ . This ensures the densities live on a common space $X_{-k}$ .

Key Results and Illustrations

Bivariate Gaussian Example: The paper provides a fully worked example with two models (A and B) that are observationally equivalent (identical joint Gaussian density and score field).
- Model A: $Y = \beta X + \epsilon_Y$ . Intervention on $X$ shifts the mean of $Y$ . The interventional score sensitivity $S_{X \to Y} = \beta/\sigma^2 \neq 0$ .
- Model B: $X = \gamma Y + \epsilon_X$ . Intervention on $X$ does not affect $Y$ . The interventional score sensitivity $S_{X \to Y} = 0$ .
- This demonstrates that while the static observational geometry is identical, the interventional score derivatives diverge, correctly identifying the causal direction only when structural assumptions are supplied.
Classical Designs Reinterpreted:
- Randomized Controlled Trials (RCTs): Randomization ensures the treatment $T$ is independent of covariates $Z$ . Geometrically, this manifests as a block-separation in the joint score field $\psi_{T,Z}$ , where the covariate component is invariant to changes in $T$ . The paper clarifies that randomization guarantees a "fair starting line" (independence of assignment) but does not geometrically flatten the treatment-outcome relationship.
- Instrumental Variables (IV): The IV estimand is restated as a ratio of directional derivatives of conditional mean functions ( $D_{v_Z} m_Y / D_{v_Z} m_T$ ), rather than a ratio of raw variables.
- Score-Based Generative Models: The paper explicitly warns that training generative models to estimate the observational score $\nabla \log p(x)$ does not yield causal structure. Without interventional data to estimate the response field, the model remains on the "associational rung" of causality.

Significance and Claims
The paper claims to provide a "corrected geometric dictionary" for causal inference that strictly separates what can be learned from observational data versus what requires structural or experimental input.

Modesty of Claims: The author explicitly states the framework does not manufacture identification power beyond existing assumptions. It does not offer new asymptotic results for meta-learners or double machine learning, nor does it solve the identification problem where structural assumptions are absent.
Scope: The framework is deliberately narrow, focusing on the "interventional rung" of Pearl's Ladder. It leaves the "counterfactual rung" (unit-level cross-world structure) for future work.
Terminological Clarity: The paper distinguishes its use of "score" (gradient of log-density) from the "score" used in semiparametric efficiency and double machine learning (estimating equations), preventing conflation of distinct mathematical objects.

In summary, the paper argues that causal inference requires moving from a static geometry of $p(x)$ to a dynamic geometry of interventional families. The "causal" information is not hidden in the curvature of the observational manifold but is encoded in the deformation of the score field under explicit, structurally defined interventions.