Ollivier–Ricci Curvature in Discrete Geometries

Updated 21 January 2026

Ollivier–Ricci curvature is a transport-based geometric measure that quantifies curvature on discrete spaces and graphs using optimal transport between local probability measures.
It is computed via the 1-Wasserstein distance through linear programming and efficient approximations, incorporating idleness parameters to model random walks.
The concept is pivotal in network science, machine learning, and quantum information, where it aids in understanding clustering, structural fragility, and manifold recovery.

Ollivier–Ricci Curvature

Ollivier–Ricci curvature provides a rigorous, transport-based framework for defining Ricci curvature on discrete metric measure spaces, particularly graphs and hypergraphs. Formulated originally by Y. Ollivier, this notion characterizes the local geometry of a network in terms of optimal transport between probability measures associated to neighborhoods. The curvature reflects how the average distance between “neighborhoods” of two points compares to the distance between the points themselves, thus extending the classical intuition from Riemannian geometry to the discrete and combinatorial setting. Ollivier–Ricci curvature has deep connections to optimal transport theory, Markov chains, spectral properties, geometric group theory, and serves as a geometric invariant in machine learning, data analysis, and network science.

1. Formal Definition and Fundamental Structure

Let $(X, d)$ be a metric space (often a graph $G=(V,E)$ with shortest-path metric $d(x,y)$ ). At each $x$ , a probability measure $\mu_x$ —typically supported on immediate neighbors—is defined. The 1-Wasserstein (earth-mover) distance between $\mu_x$ and $\mu_y$ is

$W_1(\mu_x, \mu_y) = \inf_{\pi \in \Pi(\mu_x, \mu_y)} \int_{X \times X} d(u,v) \, d\pi(u,v),$

where $\Pi(\mu_x, \mu_y)$ denotes the set of couplings (joint measures) with marginals $\mu_x$ and $G=(V,E)$ 0. The Ollivier–Ricci curvature along $G=(V,E)$ 1 is

$G=(V,E)$ 2

For graphs, a common choice is the uniform measure on neighbors: $G=(V,E)$ 3 An “idleness” parameter $G=(V,E)$ 4 allows for lazy random walks: $G=(V,E)$ 5 which interpolates Dirac ( $G=(V,E)$ 6) and uniform walk ( $G=(V,E)$ 7) cases (Wee et al., 2020, Pal et al., 2017). By Kantorovich duality, $G=(V,E)$ 8 can be computed as a variational supremum over 1-Lipschitz functions: $G=(V,E)$ 9

2. Computation and Structural Properties

The optimal transport problem underlying $d(x,y)$ 0 is solved as a linear program; for an edge $d(x,y)$ 1, this involves measures supported on $d(x,y)$ 2 and $d(x,y)$ 3 and a cost matrix of $d(x,y)$ 4 for $d(x,y)$ 5, $d(x,y)$ 6. For regular graphs ( $d(x,y)$ 7-regular, $d(x,y)$ 8), explicit assignment formulas and symmetries permit further optimization (Hehl, 2024). The idleness function $d(x,y)$ 9 is concave and piecewise linear with at most three segments (two for regular graphs); the Lin–Lu–Yau derivative

$x$ 0

proves particularly useful for analytic purposes (Bourne et al., 2017). The metric can be generalized to weighted, directed graphs and extended to random walks with arbitrary transition kernels (Fathi et al., 2022).

For practical computation, exact $x$ 1 is $x$ 2 per edge in naïve approaches, where $x$ 3. Fast combinatorial lower bounds (Jost–Liu), efficient assignment reductions, and approximation schemes (entropic regularization, sliced Wassserstein) mitigate cost for large-scale networks (Kang et al., 2024, Pal et al., 2017). Approximations based on Jaccard similarity (JC, gJC) enable nearly linear-time proxies on massive graphs, with gJC matching OR curvature in key stochastic-graph regimes (Pal et al., 2017).

3. Theoretical Significance and Limit Transitions

Ollivier–Ricci curvature provides a discrete analogue converging, under suitable limits, to the classical Ricci curvature of smooth Riemannian manifolds. In random geometric graphs sampled from an $x$ 4-manifold $x$ 5, Van der Hoorn, Trillos, and collaborators, and Hickok–Blumberg and team established rigorous $x$ 6 and high-probability convergence: $x$ 7 with graph parameters (connectivity radius $x$ 8, measure radius $x$ 9) tuned for scaling windows (Hoorn et al., 2020, Hoorn et al., 2020, Hickok et al., 6 Oct 2025). Scalar curvature can be discretized as a weighted sum over edgewise Ollivier–Ricci curvatures, converging (with explicit scaling) to the smooth scalar curvature (Hickok et al., 6 Oct 2025).

On discrete structures, such as hypergraphs or directed hypergraphs, the curvature is defined by suitably aggregated optimal transport between local probability measures (extensions in ORCHID, equal-edge/nodes walks, barycentric aggregation, and extensions to directed hyperedges using in-/out-measures) (Coupette et al., 2022, Eidi et al., 2019). All generalizations reduce to the classical case for $\mu_x$ 0-uniform graphs and appropriate measure choices.

4. Geometric Interpretation and Network Invariants

Ollivier–Ricci curvature quantifies how the neighborhoods of nodes $\mu_x$ 1 overlap compared to their separation:

$\mu_x$ 2: neighborhoods are tightly coupled—indicative of significant clustering and redundancy, as in cliques or dense motifs.
$\mu_x$ 3: neighborhoods are far apart—signaling tree-like or bottleneck structures, local negative curvature, and potential for congestion or structural fragility (Ni et al., 2015, García et al., 2024).
For $\mu_x$ 4,

$\mu_x$ 5

with tightness in trees (no triangles), and improvement by the presence of shared neighbors (triangles), quantitatively linking OR curvature to clustering coefficients (Jost et al., 2011).

OR curvature relates to spectral properties: a lower bound on $\mu_x$ 6 is equivalent to exponential decay of Lipschitz seminorms under the heat semigroup generated by the Laplacian, as in classical curvature—dimension inequalities of Bakry–Émery theory (Münch et al., 2017). The discrete Ricci flow on graphs, evolving edge weights/metrics according to $\mu_x$ 7, preserves this geometric invariant and is well-posed for a wide class of graphs (Fathi et al., 2022, Münch et al., 2017, Torbati et al., 1 Jan 2025).

5. Extensions: Persistent, Quantum, and Algorithmic Frameworks

The concept of persistent Ollivier–Ricci curvature (OPRC) is structured for attributed/molecular graphs. Using persistent filtration analogous to persistent homology, one computes $\mu_x$ 8 across a nested graph sequence and summarizes curvature-value trajectories into statistical descriptors (moments, variation, persistence); these invariants provide scalable multi-scale features in drug design, outperforming traditional molecular fingerprints in binding affinity prediction (Wee et al., 2020).

Ollivier–Ricci curvature has been generalized to non-commutative (quantum) spaces as coarse Ricci curvature of quantum channels. The contraction property for quantum Wasserstein distances leads to functional inequalities (Poincaré, spectral gap) and bounds for quantum dynamical semigroups, with explicit examples (Gibbs samplers, beam-splitters, Pauli channels) (Gao et al., 2021).

Quantum algorithms allow provably exponential speedups for OR curvature estimation on tree metrics and small-neighbor graphs (using QSVT, block-encoding, and amplitude estimation frameworks), facilitating applications in combinatorial quantum gravity and geometric ML (Nghiem et al., 10 Dec 2025).

6. Applications and Interpretative Insights

Ollivier–Ricci curvature is utilized across several domains:

Network Science: Negative curvature identifies critical bottlenecks, backbone links, or systemic fragility (e.g., Internet topology, financial networks, market crises), while positive curvature signals redundancy and local resilience (Ni et al., 2015, García et al., 2024).
Machine Learning & Data Analysis: As a feature in graph-based ML, OPRC produces state-of-the-art performance in structure-based drug design (Wee et al., 2020). In GNNs, the sign of $\mu_x$ 9 predicts over-smoothing ( $\mu_x$ 0) and over-squashing ( $\mu_x$ 1), motivating rewiring strategies (Batch Ollivier–Ricci Flow, BORF) that improve deep GNN performance (Nguyen et al., 2022).
Manifold Learning & Geometry Processing: Discrete OR curvature enables intrinsic geometry recovery and informs manifold hypothesis tests in network embeddings, persistent TDA, and community detection, with rigorous continuum convergence results (Hoorn et al., 2020, Hickok et al., 6 Oct 2025, Torbati et al., 1 Jan 2025).
Quantum Information: Coarse Ricci curvature for quantum channels connects transport contraction to spectral gap and recovery bounds in quantum dynamical systems (Gao et al., 2021).

7. Methodological and Practical Considerations

From a computational perspective, practical estimation of Ollivier–Ricci curvature is constrained by the cost of small-scale optimal transport. For practical graphs:

Efficient lower bounds and proxies (gJC, JC) allow extension to $\mu_x$ 2–scale graphs with moderate loss of geometric fidelity (Pal et al., 2017, Kang et al., 2024).
For regular graphs, explicit assignment-based algorithms reduce the problem to linear assignment (Hungarian or Bertsekas' algorithm) (Hehl, 2024).
For hypergraphs, the ORCHID framework generalizes curvature computation using barycentric or mean-pair aggregation; the theoretical properties (Bonnet–Myers bounds, diameter constraints) are retained (Coupette et al., 2022, Eidi et al., 2019).
For molecular and filtration-based applications, feature summarization pipelines (with up to $\mu_x$ 3–dimensional descriptors) provide robust high-throughput featurization (Wee et al., 2020).
Quantum algorithms yield polylogarithmic-depth algorithms in tree-structured or constant-degree contexts, with resource requirements polynomial in neighborhood sizes (Nghiem et al., 10 Dec 2025).

Limitations remain in scaling to highly irregular graphs, evaluating long-range (nonlocal) curvature, or generalizing beyond pairwise metric structures. Various open problems pertain to the classification of bone-idle (flat) edges/graphs, theoretical properties of Ricci flows on complex networks, and extensions to broader classes of weighted, directed, or non-local structures (Bourne et al., 2017, Hehl, 2024, Coupette et al., 2022).

References (arXiv IDs):