Graph Ricci Curvature and Causal Inference Reliability

10 concepts3 papers1 proposal

Causal inference on networks is notoriously difficult. When units are connected—such as individuals in a social network or proteins in a biological pathway—the assumption of independence vanishes. A treatment applied to one node often "spills over" to its neighbors, a phenomenon known as interference or peer effects.

Recent research has identified that the difficulty of estimating these effects is not just a matter of sample size, but a fundamental property of the network's geometry. Specifically, the local Ricci curvature of a graph serves as a geometric signature that determines the reliability of causal estimations.

The Geometry of Information Flow

In differential geometry, Ricci curvature describes the degree to which a space deviates from being "flat." When applied to graphs, discrete Ricci curvature measures the connectivity and redundancy between nodes.

  • Positive Curvature: These regions are dense and highly interconnected, often characterized by many triangles and redundant paths. In these areas, information (or "influence") flows easily through multiple channels.
  • Negative Curvature: These regions are sparse and tree-like, often containing "bridges" or bottlenecks. Here, information flow is constrained to single paths.

Farzam et al. theoretically and empirically demonstrate that this geometry directly impacts causal inference. They found that regions with high negative curvature significantly hinder the accuracy of Individual Treatment Effect (ITE) estimations. Conversely, regions with positive curvature facilitate more reliable inference.

Why Curvature Predicts Reliability

The link between geometry and causality is rooted in the concepts of system and distributional robustness. As Farzam et al. argue, Ricci curvature acts as a bridge between a network's ability to recover from perturbations and the invariance of its causal parameters across different environments.

In positively curved regions, the network is more "robust." The authors show that in these dense regions, the correct causal model is more easily identifiable because any alternative, incorrect model would require significantly higher entropy to fit the observed data. In contrast, the bottlenecks found in negatively curved regions (like the thin branches of a tree) make the system fragile. In these areas, small errors in data or model specification are amplified, leading to high estimation errors.

This geometric perspective addresses a gap noted by Pretschner, who suggested that incorporating structural information like geodesic distances could refine the accuracy of causal attributions. Ricci curvature provides a sophisticated way to quantify this structural information.

Flattening the Network for Better Inference

If negative curvature is the problem, the solution is to change the geometry. Farzam et al. propose a preprocessing technique using Ricci flow.

Ricci flow is a process that evolves the weights of the graph edges to "smooth out" the geometry. By iteratively adjusting the network, the authors can "flatten" the graph, reducing the intensity of negative curvature regions. When Graph Neural Networks (GNNs) are trained on these geometrically modified networks, their ability to estimate causal effects improves significantly.

Essentially, by preprocessing the network to minimize negative curvature, the model is less likely to be misled by the structural bottlenecks that typically obscure true causal mechanisms. This approach transforms the challenge of causal identification from a purely statistical problem into a topological one, treating the reliability of inference as an intrinsic property of the graph itself.

Key Foundational Work

Curvature and Causal Inference in Network Data (2023) — contributed 5 concept elements to this direction.
From Checking to Inference: Actual Causality Computations as Optimization Problems (2020) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_graph_geometry_and_causal_inference_accfcbac,
  title = {graph geometry and causal inference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-geometry-and-causal-inference/rdr_accfcbac38fc479ea86a38074a84d5b8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_graph_geometry_and_causal_inference_accfcbac_v1,
  title = {graph geometry and causal inference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-geometry-and-causal-inference/rdr_accfcbac38fc479ea86a38074a84d5b8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake