📊 statistics

A Mathematical Framework for Topological Causal Data Analysis

This paper introduces Topological Causal Data Analysis (TCDA), a framework that applies topological methods to generate stable, shape-sensitive summaries of causal effects for structured outcomes at both individual and distribution levels, while clarifying the conditions for identification, consistency, and the limited role of topology in causal discovery.

Original authors: Hugo Gobato Souto, Ioannis Diamantis

Published 2026-07-31

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Hugo Gobato Souto, Ioannis Diamantis

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: "Did this specific action cause that specific result?" In the world of science, this is called causal analysis. Usually, detectives look at simple numbers, like "Did the medicine lower the patient's fever by 5 degrees?" or "Did the new policy increase sales by 10%?" This works great when the result is a single number. But what if the result is a complex, squiggly shape? What if the "fever" is actually a 3D tumor growing in a weird way, or the "sales" are a tangled web of social connections? In these cases, you can't just subtract one number from another to see the difference. You need a way to measure the shape of the change, not just the size.

This is where Topological Data Analysis (TDA) comes in. Think of TDA as a pair of magical glasses that let you see the "skeleton" of data. Instead of getting lost in the messy details, these glasses highlight the big structural features: how many separate islands are there? Are there any loops or tunnels? How many holes are in the doughnut? It's like looking at a cloud and counting the holes in it, rather than measuring how much water it weighs. Scientists use this to study everything from brain networks to galaxy clusters. But here's the catch: just because you can see a cool shape doesn't mean you know what caused it. You still need the rules of the game (causal assumptions) to know if a treatment actually changed the shape, or if the shape just happened to be that way.

The Big Idea: A New Toolkit for Shape-Shifting Science

This paper introduces a new framework called Topological Causal Data Analysis (TCDA). Think of it as a master blueprint for scientists who want to ask, "Did this treatment change the shape of the outcome?" The authors, Hugo Gobato Souto and Ioannis Diamantis, realized that mixing "causal science" (figuring out cause and effect) with "topology" (studying shapes) was a bit like trying to bake a cake while wearing oven mitts and a chef's hat at the same time—it gets confusing. They wanted to separate the ingredients clearly so no one gets burned.

Their main finding is that you have to keep four distinct layers of the problem separate, like layers in a cake, to avoid mixing up your questions:

The Observation Space: What are we looking at? (A tumor, a network, a cloud of points?)
The Causal Model: What are the rules of the game? (Did we flip a coin to assign treatment? Did we control for other factors?)
The Topological Representation: How do we turn the messy data into a shape? (Do we look for loops? Holes? Connected islands?)
The Causal Query: What exactly are we asking? (Did the shape change? Did the number of holes increase?)

The paper argues that you cannot just throw a shape at a causal model and expect an answer. You have to define the shape first, then ask the causal question.

The Two Ways to Measure Shape Changes

The paper discovers that there are two very different ways to measure how a treatment changes a shape, and they don't always agree. The authors use a playful metaphor to explain this: The "Individual vs. The Crowd."

1. The "Individual" Approach (Outcome-Level TCDA)
Imagine you have a bag of marbles. You treat half of them with a magic potion.

The Method: You take each marble, look at its shape, measure its "topological features" (like how many bumps it has), and then average those measurements.
The Result: You get the average change in shape for a single marble.
The Catch: This works well if you care about how the treatment affects an individual object. But if the treatment changes the arrangement of the marbles (like making them clump together), this method might miss it because it's looking at them one by one.

2. The "Crowd" Approach (Distribution-Level TCDA)
Now, imagine you don't look at the marbles one by one. Instead, you look at the entire bag as a single object.

The Method: You take the whole group of marbles, see how they are arranged, and measure the shape of the entire group. Then you compare the "shape of the group" before and after the potion.
The Result: You might find that the potion didn't change the shape of any single marble, but it made the whole group split into two separate clusters.
The Catch: This method is great for seeing big structural changes in a population, but it's harder to calculate because you have to figure out the shape of the whole crowd first.

The paper proves mathematically that these two approaches are not the same. Sometimes, the "Individual" approach says "No change," while the "Crowd" approach says "Huge change!" (For example, if a treatment makes a single cluster of data split into two separate islands, the average shape of an individual might look the same, but the shape of the whole group has definitely changed).

What the Paper Rules Out (The "No Magic" List)

It is very important to know what this paper says you cannot do. The authors are very clear about the limits of their new toolkit:

Topology is not a time machine: You cannot look at a shape and magically know what caused it. If you see a loop in your data, topology doesn't tell you if Treatment A caused it or if it was just a coincidence. You still need to set up a proper experiment or use strong assumptions (like randomization) to know the cause.
Topology doesn't fix bad data: If your data is messy or you haven't controlled for the right factors (confounders), topology won't clean it up. It's a magnifying glass, not a magic eraser.
It doesn't always identify the full story: The paper shows that sometimes you can identify a "coarse" topological effect (like "the number of holes changed") without knowing the full details of the outcome. But you can't use this to identify the entire causal law if the shape representation is too simple.

How Sure Are They?

The authors are very confident in the math they built. They didn't just guess; they wrote down strict rules (theorems) that prove:

If you follow their four-layer framework, your questions will be clear.
If you use specific mathematical tools (like "distance-to-measure" or "persistence diagrams"), you can prove that small errors in your data won't blow up your results. They showed that if your estimate of the data is close to the truth, your estimate of the "shape change" will also be close to the truth.
They proved that the "Individual" and "Crowd" approaches only give the same answer in very specific, rare cases. In most real-world scenarios, they will give different answers, and you have to choose which one fits your scientific question.

Why Should You Care?

This framework is like a new set of instructions for scientists studying complex things.

For a doctor: If you are studying a tumor, you might not just care if it got smaller (a number). You might care if it got more "spiky" or if it developed new tunnels. This framework helps you ask, "Did the drug change the tumor's structure?"
For a climate scientist: If you are studying weather patterns, you might care if the storm system broke into two separate parts. This framework helps you measure that change.
For a network researcher: If you are studying how people connect, you might care if the network developed a giant "loop" of friends.

The paper doesn't solve every mystery, and it doesn't replace the need for good experiments. But it gives scientists a clear, safe way to ask questions about the shape of the world, ensuring they don't get confused between the shape of a single object and the shape of the whole crowd. It turns a messy, confusing problem into a structured, solvable one.

Technical Summary: A Mathematical Framework for Topological Causal Data Analysis

Problem Statement

Modern scientific outcomes often manifest as structured objects—such as images, point clouds, networks, and spatial fields—rather than scalar values. In such contexts, the standard causal contrast $Y_1 - Y_0$ is frequently undefined or scientifically inadequate. Reducing these complex structures to a single scalar summary may discard critical geometric or topological information, while simple subtraction fails to capture changes in shape, connectivity, or clustering.

While Topological Data Analysis (TDA) provides multiscale summaries (e.g., persistence diagrams, landscapes) that are stable and shape-sensitive, these tools lack an inherent causal interpretation. A topological feature describes observational geometry but does not account for confounding or treatment assignment. Conversely, standard causal inference frameworks are typically designed for scalar or vector-valued outcomes. There is a need for a rigorous mathematical architecture that integrates causal identification with topological representation without conflating the distinct roles of intervention, identification, and feature extraction.

Methodology: The TCDA Framework

The paper introduces Topological Causal Data Analysis (TCDA), a framework defined by the tuple $P_{TCDA} = (S, M(G), T, C)$ . This architecture explicitly separates four layers:

Observation Space ( $S$ ): The measurable space of observed data $(X, A, Y)$ .
Causal-Model Class ( $M(G)$ ): The set of causal models compatible with a graph $G$ , determining potential outcomes $Y^a$ .
Topological Representation ( $T$ ): A measurable map transforming outcomes or laws into a metric or Banach space.
Causal Query ( $C$ ): The specific target of inference defined on the causal model.

The framework distinguishes between two primary levels of analysis, which generally do not commute:

1. Outcome-Level TCDA

In this branch, topology is applied to individual potential outcomes before averaging.

Representation: Let $T_{out}: Y \to B$ be a measurable map into a separable Banach space $B$ (e.g., persistence landscapes, silhouettes).
Target: The Topological Average Treatment Effect (TATE) is defined as:
$TATE_{out} = E[T_{out}(Y^1)] - E[T_{out}(Y^0)]$
Identification: Under standard assumptions (Consistency, Conditional Exchangeability, Strict Positivity), the outcome-level effect is identified via standardization or inverse-probability weighting (IPW). The paper extends these results to Banach-space-valued outcomes, establishing conditions for Bochner integrability and deriving doubly robust estimators with product-rate bounds for nuisance functions.
Vectorization-Free Effects: For cases where vectorization is avoided, the framework considers metric-space estimands like Fréchet means or geometric medians of persistence diagrams, relying on the metric geometry of the diagram space (e.g., Wasserstein distance).

2. Distribution-Level TCDA

In this branch, topology is applied to the interventional outcome laws themselves.

Representation: Let $T_{dist}: \mathcal{P}^*(Y) \to Z_{dist}$ be a map acting on probability laws (e.g., applying persistent homology to the support or distance-to-measure function of the law).
Target: The distribution-level effect is:
$\Delta_{dist} = T_{dist}(P^1_Y) - T_{dist}(P^0_Y)$
(or a scalar discrepancy $d_{dist}(T_{dist}(P^1_Y), T_{dist}(P^0_Y))$ ).
Identification: The interventional laws $P^a_Y$ are identified via the causal $g$ -formula. The topological representation is applied after this identification step.
Non-Commutation: The paper demonstrates that outcome-level and distribution-level contrasts generally differ. They agree if and only if $T_{dist} - A$ is constant on the relevant class of laws (where $A$ is the mean functional). Non-affine representations (e.g., those sensitive to clustering or connectivity) create a fundamental divergence between averaging topological summaries and summarizing averaged laws.

Stability and Plug-in Guarantees

The framework establishes that if the topological representation is Lipschitz continuous with respect to a specific diagram metric (e.g., bottleneck or Wasserstein distance), the causal contrast inherits this stability.

Outcome Level: Stability transfers from the outcome space to the TATE via the coupling of potential outcomes.
Distribution Level: Stability transfers from the space of probability laws (e.g., $W_2$ distance) to the distribution-level effect. For Distance-to-Measure (DTM) persistence, the error in the estimated effect is bounded by the $W_2$ error of the estimated interventional laws, scaled by the DTM parameter $m$ .

Topological Ignorability and Discovery Limits

Target-Specific Ignorability: Building on Saki and Faghihi [32], the paper formalizes conditional topological ignorability. This is a weaker assumption than weak conditional exchangeability, sufficient to identify covariate-standardized topological effects even when the full interventional laws are not identified, provided the representation is non-injective and distinct laws lie in the same fiber.
Causal Discovery: The paper clarifies the limits of topology in causal discovery. While topological summaries can distinguish between restricted model classes (e.g., separating a latent circular mechanism from functional dependencies based on support topology), they cannot, by themselves, orient causal edges or identify interventions without additional assumptions (e.g., temporal order, independent noise).

Key Contributions

Unified Architecture: A formal separation of observation, causal models, topological representations, and queries, preventing the conflation of scientific questions regarding interventions, identification, and feature relevance.
Banach-Space Formulation: A rigorous treatment of outcome-level TCDA for Banach-space-valued summaries, including identification, doubly robust estimation, and conditions for Bochner integrability.
Characterization of Non-Commutation: A theoretical characterization of when outcome-level and distribution-level effects agree, proving that non-affine topological transformations generally lead to distinct causal targets.
Stability Transfer: Derivation of stability bounds for causal contrasts based on the Lipschitz constants of topological representations, specifically linking $W_2$ -errors in law estimation to errors in DTM-based causal effects.
Clarification of Discovery Limits: A formalization of when observational topology can assist in causal discovery (via separation of model classes) and explicit statements on its inability to replace causal assumptions for structure identification.

Results and Significance

The paper does not propose new estimation algorithms for specific datasets but rather provides the mathematical infrastructure to justify and analyze existing and future methods.

Significance: The framework clarifies that topology does not define interventions; it supplies stable, shape-sensitive summaries after causal assumptions are specified. It highlights that the order of operations (topological transformation vs. population aggregation) defines the causal target.
Practical Implication: Researchers must explicitly choose whether their scientific question concerns the average change in individual shapes (outcome-level) or the change in the geometry of the population distribution (distribution-level), as these are distinct targets that may yield different conclusions.
Robustness: The framework shows that while topological ignorability can allow for the identification of coarse effects under weaker assumptions, it does not provide generic robustness to hidden confounding for marginal effects or full law identification.

In conclusion, TCDA places topological data analysis within the standard causal inference sequence (target definition $\to$ assumptions $\to$ identification $\to$ estimation), ensuring that topological summaries are interpreted with appropriate causal rigor.