📊 statistics

A Mathematical Framework for Topological Causal Data Analysis

This paper introduces Topological Causal Data Analysis (TCDA), a framework that integrates stable topological summaries with causal inference to analyze structured outcomes like images and networks, providing identification, estimation, and consistency results for both individual and distribution-level causal effects while clarifying the limits of topology in causal discovery.

Original authors: Hugo Gobato Souto, Ioannis Diamantis

Published 2026-09-18

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Hugo Gobato Souto, Ioannis Diamantis

Original paper licensed under CC BY 4.0 (https://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

In the world of modern science, researchers often face a peculiar problem: the things they want to study are too complex to be reduced to a single number. When a doctor treats a tumor, the result is not just a smaller mass; it is a change in shape, a rearrangement of internal tunnels, or a shift in how the tissue connects to itself. Similarly, a climate scientist might look at a weather map not just to see if it got hotter, but to understand if the storm patterns have become more fragmented or if new loops of circulation have formed. Traditional statistics struggle here because they are built to compare averages, like the difference between two heights or two weights. But you cannot simply subtract one shape from another to find a meaningful answer, nor can you capture the essence of a network by averaging its connections. To understand these changes, scientists need a way to measure the geometry and the "holes" within data, a field known as topological data analysis. This approach treats data like a physical object that can be stretched and examined at different scales to see which features persist and which disappear.

A new framework called Topological Causal Data Analysis, developed by Hugo Gobato Souto and Ioannis Diamantis, brings this geometric perspective into the rigorous world of cause and effect. For decades, scientists have used causal inference to determine if a treatment actually causes an outcome, separating the effect of a drug from the natural history of a disease. However, these methods were designed for simple numbers. The researchers realized that to study complex outcomes like brain networks or molecular structures, they needed a new mathematical architecture that respects both the shape of the data and the logic of causality. Their work does not invent a new way to prove that a drug works; rather, it provides a precise set of rules for how to measure the shape of the change once the causal link has been established. They show that topology, the study of shape and connectivity, is a powerful tool for summarizing complex data, but it must be applied carefully so that it does not confuse the geometry of the data with the cause of the change.

The core of the paper is a four-part structure that keeps the different layers of a scientific question separate. First, there is the observation space, which is simply the raw data, such as a 3D image of a tumor or a map of a brain. Second is the causal model, which defines what the researchers believe caused the change, such as a specific drug or an environmental factor. Third is the topological representation, a method for turning that complex shape into a mathematical summary that captures its essential features, like the number of loops or voids. Finally, there is the causal query, which asks the specific question, such as "Did the treatment change the number of tunnels in the tumor?" By keeping these four layers distinct, the framework prevents scientists from accidentally mixing up the shape of the data with the cause of the change. The authors demonstrate that topology does not define the intervention itself; it only provides a stable, shape-sensitive way to describe the result after the causal assumptions have been made.

The researchers identified two fundamentally different ways to apply this framework, and they found that these two approaches often lead to different answers. The first approach, which they call outcome-level analysis, looks at each individual patient or object, measures its shape, and then averages those measurements across the group. Imagine measuring the shape of every single tumor in a study and then finding the average shape change. The second approach, distribution-level analysis, takes a different path. It first combines all the data to form a complete picture of the population's behavior under treatment, creating a single "law" that describes the group, and then measures the shape of that entire group picture. The paper proves that these two methods are not interchangeable. In many cases, the average of the individual shapes is not the same as the shape of the average group. For instance, a treatment might leave the average size of a tumor unchanged, but it could cause the population to split into two distinct groups: some tumors shrinking into tight knots and others expanding into loose clouds. The first method might miss this split entirely, while the second method would clearly detect the new, fragmented structure.

A significant portion of the work is dedicated to ensuring that these measurements are reliable. The authors show that if the mathematical tools used to describe the shapes are stable—meaning that a tiny error in the data does not cause a huge jump in the result—then the causal conclusions drawn from them are also stable. They provide specific mathematical guarantees for several common ways of measuring shapes, such as those based on the distance between points or the density of the data. This is crucial because real-world data is always noisy. The framework ensures that if a scientist uses a robust method to measure the shape of a brain network, they can trust that a detected change in the network's loops is a real effect of the treatment and not just a glitch in the measurement.

The paper also addresses a common misconception: that looking at the shape of data can automatically reveal the cause of a relationship. The authors are clear that this is not true. While topological patterns can help scientists diagnose whether a model is missing something—for example, by revealing a circular pattern in the data that a straight-line model failed to capture—topology alone cannot determine which variable caused the other. A loop in a graph does not tell you which way the arrow of time points. To find the cause, scientists still need to rely on established assumptions, such as randomization or specific knowledge about how the system works. The framework simply adds a new, powerful lens for viewing the results once those assumptions are in place.

Finally, the researchers explore a scenario where the topological summary is so coarse that it might hide some details, yet still allow for a valid causal conclusion. They show that under certain conditions, a scientist can identify the effect of a treatment on a specific topological feature without needing to know the full, detailed distribution of the data. This is a subtle but important finding, suggesting that sometimes a simplified view of the shape is enough to answer a specific scientific question, even if the full picture remains too complex to map. The work concludes by placing topology firmly within the standard scientific process of defining a target, making assumptions, and estimating effects. It does not replace the need for careful causal reasoning, but it offers a rigorous way to ask and answer questions about the shape of the world, from the folds of a protein to the structure of a climate system.

Technical Summary: A Mathematical Framework for Topological Causal Data Analysis

Problem Statement
Modern causal inference frequently encounters outcomes that are structured objects—such as images, point clouds, networks, and spatial fields—rather than scalar values. For such outcomes, the standard definition of the Average Treatment Effect (ATE), $E[Y_1] - E[Y_0]$ , is often scientifically inadequate or mathematically undefined because subtraction is not well-defined for shapes or graphs. Furthermore, reducing these complex structures to a single scalar summary risks discarding the very geometric or topological features of interest. While Topological Data Analysis (TDA) offers multiscale summaries (e.g., persistence diagrams, landscapes) that capture shape, these tools lack an inherent causal interpretation; a topological feature may describe observational geometry without corresponding to a causal intervention. The paper addresses the need for a rigorous mathematical framework that integrates causal assumptions with topological representations to define, identify, and estimate causal effects on structured outcomes.

Methodology and Framework
The authors propose Topological Causal Data Analysis (TCDA), a framework defined by the tuple $P_{TCDA} = (S, M(G), T, C)$ , which explicitly separates four distinct layers:

Observation Space ( $S$ ): The space of observed data.
Causal-Model Class ( $M(G)$ ): The class of causal models compatible with a given causal graph $G$ .
Topological Representation ( $T$ ): A measurable map transforming outcomes or probability laws into a topological summary space.
Causal Query ( $C$ ): The specific causal contrast of interest.

The framework distinguishes between two primary levels of analysis:

Outcome-Level TCDA: Applies the topological representation $T_{out}$ to individual potential outcomes $Y^a$ to obtain $Z^a = T_{out}(Y^a)$ , and then computes the effect as the difference in expectations: $TATE_{out} = E[Z^1] - E[Z^0]$ . This branch builds on existing functional causal machinery, requiring the outcomes to be Bochner integrable in a separable Banach space.
Distribution-Level TCDA: First forms the interventional law $P^a_Y = \mathcal{L}(Y^a)$ and then applies a representation $T_{dist}$ directly to the law: $\Delta_{dist} = T_{dist}(P^1_Y) - T_{dist}(P^0_Y)$ . This approach captures changes in population-level geometry (e.g., clustering, connectivity) that may not be reflected in the mean of individual outcomes.

Key Contributions and Results

Unified Architecture and Non-Commutativity:
The paper formalizes the separation of the causal target from its identification. A central theoretical result (Theorem 5.9) characterizes when outcome-level and distribution-level contrasts agree. They agree for all pairs of laws if and only if the difference between the distribution-level map and the mean functional is constant. Since most topological representations (e.g., those based on Betti numbers or support topology) are non-affine, the two levels generally define different causal targets and do not commute.
Identification and Estimation:
- Outcome-Level: Under standard assumptions (consistency, conditional exchangeability, strict positivity), the paper establishes identification for Banach-space-valued topological outcomes. It derives inverse-probability weighting and augmented (doubly robust) representations. For specific cases like power-weighted silhouettes, it cites existing results for efficient influence functions and functional weak convergence (Kim and Lee 2026).
- Distribution-Level: Identification is achieved via the standard causal g-formula applied to the interventional laws before the topological transformation. The paper provides plug-in consistency results, showing that if an estimator of the interventional law converges in a suitable metric (e.g., Wasserstein distance $W_2$ ), the resulting topological effect estimator converges, provided the topological map is Lipschitz continuous.
Stability and Error Propagation:
The authors derive stability bounds showing that if a topological representation is Lipschitz with respect to a diagram metric (e.g., bottleneck or Wasserstein distance), the causal contrast inherits this stability. Specifically, for Distance-to-Measure (DTM) persistence, the error in the estimated topological effect is bounded by the $W_2$ -error of the estimated interventional laws, scaled by the DTM parameter $m$ .
Target-Specific Ignorability:
The paper incorporates the concept of "target-specific topological ignorability" (Saki and Faghihi 2026). This condition, which requires that the topological summary of the potential outcome law equals that of the factual outcome law given covariates, can be strictly weaker than weak conditional exchangeability if the representation is non-injective. This allows for the identification of covariate-standardized topological effects without identifying the full interventional laws, though it does not generally identify marginal effects.
Limits of Causal Discovery:
The paper clarifies the role of topology in causal discovery. While observational topology can distinguish between restricted classes of causal models (e.g., separating a latent circular mechanism from functional additive noise models based on the Betti number of the support), it cannot, by itself, orient causal edges or identify interventions without additional assumptions (e.g., temporal order, independent noise asymmetry).

Significance and Claims
The paper claims to provide the first unified mathematical architecture for combining TDA with causal inference. Its primary significance lies in:

Clarifying the Causal Target: It demonstrates that the order of operations—transforming outcomes then averaging versus averaging laws then transforming—defines fundamentally different causal quantities.
Rigorous Identification: It establishes that topological summaries do not replace causal assumptions; rather, they specify the feature on which a causal contrast is evaluated after assumptions are made.
Stability Guarantees: It provides a theoretical basis for transferring stability properties from topological representations to causal estimands, ensuring that small perturbations in data or estimated laws lead to bounded changes in the causal conclusions.
Scope Definition: It modestly delineates the limits of the framework, noting that while topology aids in diagnosing residual structure and distinguishing specific model classes, it does not solve the general problem of causal discovery or provide robustness to hidden confounding without specific, often untestable, assumptions.

The work relies on standard causal identification principles (g-formula, IPW, double robustness) and extends them to Banach-space-valued and law-valued objects, leveraging existing stability theorems from TDA literature (e.g., Cohen-Steiner et al., Chazal et al.) to ensure the resulting causal estimators are well-behaved.

More like this