Ollivier-Ricci Curvature of Riemannian Manifolds and Directed Graphs with Applications to Graph Neural Networks

arXiv:2604.14211 · math.DG, cs.AI, cs.SI, math.CO · Submitted 2026-04-06 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Ollivier-Ricci Curvature of Riemannian Manifolds and Directed Graphs with Applications to Graph Neural Networks".

Tom: This thesis provides an exposition of Ollivier-Ricci curvature, a notion defined for general metric spaces such as Riemannian manifolds and graphs.

Jane: First, who's behind it and why it matters.

Title and authors: Tom: So, to start us off, let's talk about what exactly this paper is promising. The authors are building on a foundational concept called the one-Wasserstein Distance, which is all about optimal transport. They aren't just comparing two points; they are comparing entire probability distributions of mass moving from one location to another.

Jane: That’s right, Tom. Think of it like trying to move sand from one pile to another. The Wasserstein distance tells you the cheapest way to do that while keeping all the sand intact, which is a very precise measure of how far apart two locations are in terms optimal transport cost.

Lu: And then, this paper introduces Ollivier-Ricci Curvature by comparing that optimal transport cost—the one-Wasserstein distance—with the standard Euclidean distance between two points. It’s a synthetic definition that is designed to mimic the properties of classical Ricci curvature found in smooth manifolds, but it's not derived from them.

Meng: Mimicking properties is interesting, but how does this translate into something we can actually measure on graphs? Is it just a number telling us if a certain structure is "curved" or "flat"?

Lalam: It goes beyond just flat or curved, Meng. The goal of the paper is to provide a robust framework that allows localized curvature to guide our understanding, providing insights into how strongly connected or dispersed different parts of a network are from each point.

Tom: And that's where the big picture starts to form: Ollivier-Ricci Curvature gives us this localized measure, but it's not just a single number. It lets us see if moving mass from one distribution to another is cheaper than the physical distance between two points, which is what determines if the curvature is positive or negative.

Jane: This leads directly into the core idea of how we use this concept on both undirected graphs and directed graphs, which are very different things structurally.

The paper's summary: Tom: The paper starts by giving a rigorous exposition of how Ollivier-Ricci Curvature extends classical Ricci curvature to general metric spaces. This is a huge step because we're no longer restricted to smooth, continuous surfaces.

Jane: The authors show that this new definition retains key features of classical Ricci curvature, especially regarding things like controlling geodesic divergence and volume growth. It provides a way to apply the powerful mathematical tools of differential geometry to discrete settings.

Lu: They cover the work of Lin-Lu-Yau in extending this idea to undirected graphs and also explore combinatorial bounds proven by Jost and Liu, which are essential for practical application on discrete structures.

Meng: I'm interested in those combinatorial bounds, because they suggest that we can calculate this curvature without needing complex calculus tools. It feels like a practical way to make complex network analysis computationally feasible.

Lalam: That feasibility is vital for the future work of AI. We need methods that can process large datasets efficiently, and localized curvature provides exactly that kind of information—a concise measure of local connectivity or lack thereof.

Tom: And while the authors have done a lot on undirected graphs, they' also identify and address a major challenge in their new directed graphs section. The complexity comes from the asymmetry of directed edges, which is where things get messy compared to standard undirected models.

Jane: That asymmetry means that moving mass from point X to point Y might not be the same as moving it back, so our standard assumptions about symmetry break down when working with directed paths and cycles.

The paper's improvements: Tom: So, we've covered the basics on undirected graphs; let's talk about what the paper is actually doing to push the boundaries of directed graphs. This is where things get genuinely new and exciting.

Jane: The authors are tackling that asymmetry head-on. They propose novel ideas for Ollivier-Ricci curvature on directed graphs, looking at concepts like cycles and how mass can flow through them, especially in k-cycles, which is a very complex structural element.

Lu: It's not just about applying the old formula to a modified structure; the authors are creating new theoretical bounds for directed structures that have no easy analog in the undirected world. This is where the true innovation lies.

Meng: When they talk about "filling gaps" or considering out-branching versus in-branching trees, I'm trying to picture this practically. Are we talking about optimizing traffic flow in a logistics network, perhaps?

Lalam: That's a great analogy, Meng. The way the model is designed to track the movement of mass—or information—through these directed paths allows us to see bottlenecks or areas where flow is restricted by the very structure of guiding edges.

Tom: Exactly, Lalam. We are looking at how this curvature can inform algorithms that use GNNs and network science, which are our final destination for this research.

Jane: The authors show how these localized measures—whether positive or negative —can be used in community detection and graph rewiring to find structures that are either highly connected or too sparse.

Conclusion: Tom: Well, we've covered a lot of ground today, from the foundations of optimal transport all the way through to directed graphs and applications for AI. It's clear this is a very comprehensive paper.

Jane: It really is, Tom. The authors have successfully demonstrated how a powerful geometric concept can be extended to solve problems in discrete network analysis, moving beyond the limitations of traditional methods.

Lu: I think we should all be thrilled about the implications for future research; the groundwork laid here is incredibly robust for those who want to explore new mathematical frameworks for machine learning.

Meng: The practical application, especially in areas like optimizing communication paths or improving how GNNs learn from directed data, is definitely a major win. We need these tools to handle the complexity of real-world digital information flow.

Lalam: To summarize our excitement: "Ollivier-Ricci Curvature of Riemannian Manifolds and Directed Graphs with Applications to Graph Neural Networks" offers a powerful language for understanding structure and flow that is essential for improving how AI perceives and interacts with complex, directed data structures.

math.DG, cs.AI, cs.SI, math.CO

Submitted: 2026-04-06

Updated: 2026-08-24

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 88/100

The gist: This thesis provides an exposition of Ollivier-Ricci curvature, a notion defined for general metric spaces such as Riemannian manifolds and graphs.

Key concepts

One-Wasserstein Distance: This is a measure from optimal transport that compares two probability distributions. It calculates the cheapest way to move mass from one location to another while keeping all the mass intact, providing a precise measure of separation based on transport cost.
Ollivier-Ricci Curvature: This is a synthetic definition created by comparing the optimal transport cost (one-Wasserstein distance) with the standard Euclidean distance. It mimics classical Ricci curvature properties to provide a localized measure of how strongly connected or dispersed different parts of a network are.
Directed Graphs: These graphs have asymmetric edges, meaning moving mass from point X to Y is not necessarily the same as moving it back. The paper addresses the complexity introduced by this asymmetry when applying curvature concepts to directed structures.
Graph Neural Networks (GNNs): The research aims to use localized curvature measures derived from the paper to inform algorithms in GNNs. This helps GNNs better perceive and interact with complex, directed data structures for applications like optimizing communication paths.

Terminology

Optimal Transport — 35× in this paper · explained in 5 episodes
Wasserstein Distance — 32× in this paper · explained in 4 episodes
Over-squashing — 13× in this paper · explained in 3 episodes
Over-smoothing — 12× in this paper · explained in 4 episodes
GNNs — 9× in this paper · explained in 2 episodes
Clustering — 8× in this paper · explained in 2 episodes
Graph Neural Networks (GNNs) — 3× in this paper · explained in 7 episodes
Axioms — 2× in this paper · explained in 2 episodes
Representation Learning — 2× in this paper · explained in 2 episodes
Trace — 2× in this paper · explained in 2 episodes
alpha — 2× in this paper · explained in 1 episode
Aggregation — 1× in this paper · explained in 1 episode
Core — 1× in this paper · explained in 2 episodes
Embeddings — 1× in this paper · explained in 2 episodes
Modularity — 1× in this paper · explained in 9 episodes

Summary

This thesis provides an exposition of Ollivier-Ricci curvature, a notion defined for general metric spaces such as Riemannian manifolds and graphs. It matters because this form of curvature allows for the application of geometric principles to non-derivative computation on manifolds and provides a framework to study localized curvature in complex networks, ultimately offering tools to improve algorithms in network science and graph machine learning.

The foundation of optimal transport

The mathematical basis for Ollivier-Ricci curvature is the 1-Wasserstein distance, which is rooted in the classical problem in probability theory of optimal transport. The paper distinguishes between two primary formulations used to determine the cost of moving mass from one distribution to another:

The Monge formulation, a problem where mass cannot be split, seeking an optimal transport map.
The Kantorovich relaxation, which allows for mass splitting and is formulated as a transport plan.

Through the Kantorovich-Rubinstein duality, the minimization of transport cost can be solved by maximizing the difference in expected cost of any 1-Lipschitz function. This distance metric is essential for defining curvature by comparing the distance between two points to the optimal transport cost between their associated probability measures.

Ollivier-Ricci curvature on metric spaces

Ollivier-Ricci curvature is synthetically defined to preserve properties characteristic of classical Ricci curvature. It is defined as Ric O(x, y):= 1 - W 1(m x, m y) over d(x, y). The thesis demonstrates that this notion maintains the essential geometric control found in Riemannian manifolds:

Positive Ricci curvature is characterized by the convergence of the parallel geodesics or highly connected graph clusters.
Zero Ricci curvature represents a locally flat or Euclidean surface.
Negative Ricci curvature corresponds to geodesic dispersion, such as that observed in trees.

The work connects these properties to classical results, showing that Ollivier-Ricci curvature can extend the Bonnet-Myers theorem and the Levy-Gromov isoperimetric inequality to general metric spaces, thereby bounding diameter and relating curvature to volume growth.

Extensions to undirected and directed graphs

The thesis explores how these concepts translate to discrete structures. For undirected graphs, it discusses the Lin-Lu-Yau extension, which defines curvature through a limit of alpha-lazy random walks as alpha to 1. Additionally, the work of Jost and Liu provides combinatorial bounds for graph curvature based on:

The degrees of adjacent vertices (d x and d y).
The presence of triangles, which provide positive curvature contributions.

Transitioning to directed graphs introduces a major challenge regarding the asymmetry of directed graphs and restriction of walks depending on specified directionalities. Because transport costs are not symmetric in digraphs, the paper proposes novel considerations for studying effective length in cycles and specialized bounds for in-branching and out-branching directed trees.

Applications to network science and machine learning

The final section examines how curvature-informed algorithms are applied to real-world networked data. These applications primarily address two computational tasks:

Community detection or clustering: Identifying highly-connected structure by using Ricci flow or iteratively removing the most negatively curved edge to isolate clusters.
Graph rewiring for Graph Neural Networks (GNNs): Mitigating the foundational issues of over-smoothing (where node representations become indistinguishable) and over-squashing (where information is compressed through bottlenecks).

By adjusting connections based on curvature, these methods aim to improve message-passing and enhance model expressivity in GNNs.

Improvements for AI systems

1. Directed Graph Rewiring via Asymmetric Ollivier-Ricci Curvature

Improvement: Integrate a rewiring mechanism into Graph Neural Networks (GNNs) that utilizes the proposed asymmetric directed Ollivier-Ricci curvature instead of undirected approximations. This involves using the effective length of directed cycles and in/out-degree mass transport distributions to identify topological bottlenecks.
Capability: The improved AI system can dynamically restructure directed graphs (e.g., web crawls, social follow-graphs, or gene regulatory networks) to mitigate over-squashing by adding edges that bridge directed bottlenecks and over-smoothing by removing redundant edges in highly connected strongly connected components, thereby optimizing information flow for long-range dependencies.

2. Curvature-Gated Adaptive Message Passing

Improvement: Implement an attention mechanism where the edge weights in GNN message-passing layers are conditioned on the local Ollivier-Ricci curvature value. Edges identified with highly negative curvature (bridges/bottlenecks) would trigger specialized aggregation protocols or auxiliary skip-connections.
Capability: The system can prevent signal attenuation and information loss in large-scale directed networks by automatically detecting and bypassing topological congestion points, ensuring that critical features from distant but relevant nodes are not squashed during the aggregation process.

3. Hierarchical Community Detection in Asymmetric Networks

Improvement: Replace standard stochastic block model (SBM) clustering with a curvature-informed clustering algorithm that leverages the asymmetric 1-Wasserstein distance between in-neighborhood (N in) and out-neighborhood (N out) probability measures.
Capability: The improved AI system can perform high-fidelity community detection in directed biological and citation networks, allowing it to distinguish between source communities (information providers/regulators) and sink communities (information consumers/targets), which is impossible with undirected clustering methods.

Abstract

This thesis is an exposition of Ollivier-Ricci Curvature of metric spaces as introduced by Yann Ollivier, which is based upon the 1-Wasserstein Distance and optimal transport theory. We present some of the major results and proofs that connect Ollivier-Ricci curvature with classical Ricci curvature of Riemannian manifolds, including extensions of various theoretical bounds and theorems such as Bonnet-Myers and Levy-Gromov. Then we shift to results introduced by Lin-Lu-Yau on an extension of Ollivier-Ricci curvature on graphs, as well as the work of Jost-Liu on proving various combinatorial bounds for graph Ollivier-Ricci curvature. At the end of this thesis we present novel ideas and proofs regarding extensions of these results to directed graphs, and finally applications of graph-based Ollivier-Ricci curvature to various algorithms in network science and graph machine learning.