Discrete Ricci Curvature Flow Overview

• Discrete Ricci Curvature Flow is a discrete analogue of smooth Ricci flow, evolving metric data such as edge weights, radii, or lengths to achieve prescribed curvature.
• It encompasses diverse settings including graphs with transport-based curvature, triangulated surfaces via circle packings, and piecewise-flat complexes using deficit angles.
• The method leverages convex energy functionals and stability analyses to guide convergence, with applications ranging from network alignment to geometric modeling.

Searching arXiv for the cited papers to ground the article in current records. Searching arXiv for "(Sandhu et al., 2019) discrete Ricci flow networks". Searching arXiv for "(Ge et al., 2015) hyperbolic discrete Ricci flow". Discrete Ricci curvature flow comprises a family of finite-dimensional analogues of smooth Ricci flow in which discrete metric data—typically graph edge weights, circle-packing radii, discrete conformal factors, or simplicial edge lengths—evolve proportionally to a discrete curvature. Across graphs, triangulated surfaces, and piecewise-flat manifolds, the common objective is to deform a discrete geometry toward prescribed curvature, uniformized curvature, or a geometrically informative metric. The literature does not present a single canonical discretization; instead, it develops several non-equivalent flows tied to different curvature notions, including Ollivier-type transport curvature on graphs, angle-deficit and area-normalized curvature on surfaces, and Regge- or piecewise-flat Ricci curvature on simplicial manifolds (Saucan, 2019).

1. Discrete curvature models and the scope of the flow

Discrete Ricci flow is defined relative to a chosen discrete curvature and a chosen representation of metric data. On weighted graphs, Ollivier–Ricci curvature assigns curvature to edges by comparing neighborhood measures through Wasserstein distance; the corresponding flow reweights edges according to the sign and magnitude of that curvature (Ni et al., 2018). On triangulated surfaces, combinatorial Ricci flow is usually formulated in terms of circle-packing radii or vertex conformal factors, with curvature given by angle deficit or an area-normalized variant of angle deficit (Ge et al., 2015). On piecewise-flat simplicial manifolds, curvature is concentrated on hinges or edges through deficit angles, and the flow acts directly on edge lengths (Miller et al., 2013).

A concise taxonomy is useful.

Setting	Discrete curvature	Evolving variable
Weighted graphs	Ollivier, Lin–Lu–Yau, Forman, Menger, Haantjes	edge weights
Triangulated surfaces	angle-deficit or area-normalized Gaussian curvature	radii or conformal factors
Piecewise-flat simplicial manifolds	deficit-angle, Regge, or piecewise-flat Ricci curvature	edge lengths

For surfaces, the unified framework of discrete surface Ricci flow represents several classical schemes in a common notation. It includes Thurston’s circle packing, tangential circle packing, inversive distance circle packing, discrete Yamabe, and virtual radius circle packing, with the flow written as a gradient flow for a discrete Ricci energy in the conformal variables (zhang et al., 2014). For polyhedral and graph settings, Saucan’s survey emphasizes that metric, geodesic-curvature, and combinatorial discretizations lead to distinct Ricci flows rather than a single universal construction (Saucan, 2019).

This plurality is not merely terminological. It reflects different geometric commitments: transport-based curvature probes local neighborhood coupling on graphs, angle-deficit curvature encodes discrete conformal geometry on surfaces, and Regge-type curvature models curvature concentration in piecewise-flat manifolds. A recurring misconception is therefore that “discrete Ricci flow” names one algorithm. The literature instead supports a family of flows whose analytic properties depend strongly on the underlying curvature model.

2. Ollivier-type flows on graphs and networks

For a connected weighted graph $G=(V,E,w)$, Ollivier–Ricci curvature on an edge is defined from one-step probability measures at its endpoints and the Earth-Mover distance between them. In the weighted formulation used for network control, if $\mu_i(j)=w_{ij}/d_i$ with $d_i=\sum_k w_{ik}$, then

$\kappa_{ij}=1-\frac{W_1(\mu_i,\mu_j)}{d(i,j)}.$

The open-loop flow evolves the edge weight $\mu_t(i,j)$ by

$\frac{d}{dt}\mu_t(i,j)=-\kappa_{ij}(t)\mu_t(i,j),$

so positively curved edges shrink and negatively curved edges grow (Sandhu et al., 2019).

A closely related discrete-time formulation updates graph weights by

$$w_{t+1}(u,v)=w_t(u,v)-\varepsilon\,\kappa_t(u,v)\,w_t(u,v),$$

typically followed by renormalization of total edge weight. In graph matching, this iterative reweighting is used to induce a Ricci-flow metric $d_*$, obtained as the shortest-path metric in the flowed weighted graph. That metric is then exploited through landmark distance vectors for approximate network alignment; the reported motivation is that the induced metric is comparatively insensitive to node and edge perturbations (Ni et al., 2018).

Graph flows admit several important extensions. A feedback-controlled version augments the Ollivier flow by a control term

$$\frac{d}{dt}\mu_t(i,j)=\bigl[-\kappa_{ij}(t)+\psi(\mu_t,\mu^*)\bigr]\mu_t(i,j),$$

and in the proportional-type choice $\psi=\beta_{ij}^2(t)\delta_t(i,j)$, with $\mu_i(j)=w_{ij}/d_i$0 and $\mu_i(j)=w_{ij}/d_i$1, the closed-loop system is Lyapunov stable (Sandhu et al., 2019). A prescribed-curvature version on infinite graphs evolves edge weights by

$\mu_i(j)=w_{ij}/d_i$2

with $\mu_i(j)=w_{ij}/d_i$3 the Lin–Lu–Yau curvature and $\mu_i(j)=w_{ij}/d_i$4 the target curvature; for locally finite infinite graphs, existence and uniqueness are established, and convergence is proved on graphs with girth at least $\mu_i(j)=w_{ij}/d_i$5 under two distinct hypotheses (Hua et al., 8 Jun 2026).

More abstract well-posedness results are available. For weighted graphs equipped with a metrically complete distance and sufficiently regular random walks, both discrete-time and continuous-time Ollivier–Ricci curvatures are shown to be Lipschitz in the underlying data. This yields existence and uniqueness of generalized continuous-time Ollivier–Ricci curvature flows on finite graphs through Picard–Lindelöf arguments (Fathi et al., 2022). In a separate unifying direction, piecewise-linear Ricci curvature flows on weighted graphs freeze curvature on time subintervals; for homogeneous curvatures, repeated surgeries eventually produce connected components of constant Ricci curvature (Ma et al., 21 May 2025).

Directed graphs require additional care because directed distances may be infinite. A directed Ollivier-type curvature based on outward lazy random walks leads to continuous and discrete-time Ricci flows on strongly connected directed graphs, with unique global solutions in the strongly connected case. For weakly connected directed graphs, the proposed remedy is to add artificial high-weight edges to enforce strong connectivity, run the flow, and then delete those artificial edges during core extraction (Zhao et al., 5 Dec 2025). Recent applications also use normalized discrete Ricci flow as a filtration mechanism on query–document graphs: after finite-time flow and thresholding of final weights, high-weight, negatively curved query–chunk edges are pruned before reranking in retrieval-augmented generation (Qin et al., 13 Jun 2026).

The graph literature is analytically uneven. Some formulations now have rigorous existence, uniqueness, and convergence theorems, especially in prescribed-curvature or surgery-based settings. By contrast, for the classical graph Ricci-flow metric used in alignment, a full convergence proof on arbitrary graphs remains open, even though empirical curvature uniformization is repeatedly observed (Ni et al., 2018).

3. Surface flows, circle packings, and discrete uniformization

On triangulated surfaces, discrete Ricci flow is most closely tied to discrete conformal geometry. In hyperbolic background geometry, Ge–Xu consider a closed triangulated surface $\mu_i(j)=w_{ij}/d_i$6 with circle-packing metric $\mu_i(j)=w_{ij}/d_i$7 and edge lengths determined by the hyperbolic cosine law. The classical discrete Gauss curvature is the angle deficit

$\mu_i(j)=w_{ij}/d_i$8

and the normalized discrete curvature is obtained by dividing by the area of the hyperbolic disk of radius $\mu_i(j)=w_{ij}/d_i$9,

$d_i=\sum_k w_{ik}$0

With $d_i=\sum_k w_{ik}$1 or $d_i=\sum_k w_{ik}$2, the combinatorial Ricci flow is

$d_i=\sum_k w_{ik}$3

The central theorem states that this flow converges if and only if there exists a circle-packing metric with zero curvature, and in that case the convergence is exponentially fast. A second theorem shows convergence also under initially nonpositive curvature, without requiring the a priori existence of a zero-curvature metric (Ge et al., 2015).

The unified surface Ricci-flow framework generalizes this viewpoint. There, one prescribes target curvature $d_i=\sum_k w_{ik}$4 and evolves the discrete conformal factors $d_i=\sum_k w_{ik}$5 by

$d_i=\sum_k w_{ik}$6

or the sign-reversed equivalent depending on convention. The flow is the negative gradient flow of a discrete Ricci energy

$d_i=\sum_k w_{ik}$7

whose Hessian is the Jacobian of the curvature map. This framework subsumes Thurston, tangential, and inversive-distance circle packings, as well as discrete Yamabe and virtual-radius schemes, across Euclidean, hyperbolic, and spherical background geometries (zhang et al., 2014).

Surface Ricci flow also appears in time-discretized variational form. On the “American football” $d_i=\sum_k w_{ik}$8 with conical singularities, the normalized smooth Ricci flow in conformal form is written as a gradient flow of an entropy functional $d_i=\sum_k w_{ik}$9. A discrete Morse, or minimizing-movement, scheme defines $\kappa_{ij}=1-\frac{W_1(\mu_i,\mu_j)}{d(i,j)}.$0 at each time step as the unique minimizer of

$\kappa_{ij}=1-\frac{W_1(\mu_i,\mu_j)}{d(i,j)}.$1

and the interpolants converge, after subsequence extraction, to a weak $\kappa_{ij}=1-\frac{W_1(\mu_i,\mu_j)}{d(i,j)}.$2-solution on every finite time interval (Ma et al., 2012). Although this is a discretization in time rather than in curvature alone, it is part of the broader methodology by which Ricci flows on singular surfaces are approximated in discrete form.

A more geometric specialization concerns discrete surfaces of revolution arising from circular-net theory. There, the induced metric on each quadrilateral face is diagonal, the discrete Gaussian curvature is expressed through the normal’s radial component, and the normalized flow

$\kappa_{ij}=1-\frac{W_1(\mu_i,\mu_j)}{d(i,j)}.$3

drives the metric toward constant discrete Gaussian curvature. Fixed points are precisely discrete constant-Gaussian-curvature surfaces of revolution in the integrable-systems parametrization developed in that work (Suda, 2023).

4. Piecewise-flat and simplicial flows in higher dimensions

Higher-dimensional discrete Ricci flow is usually formulated on piecewise-flat simplicial manifolds. In the Regge-calculus approach of “Simplicial Ricci Flow,” a $\kappa_{ij}=1-\frac{W_1(\mu_i,\mu_j)}{d(i,j)}.$4-dimensional simplicial complex $\kappa_{ij}=1-\frac{W_1(\mu_i,\mu_j)}{d(i,j)}.$5 is paired with its circumcentric dual $\kappa_{ij}=1-\frac{W_1(\mu_i,\mu_j)}{d(i,j)}.$6. Curvature is concentrated on codimension-$\kappa_{ij}=1-\frac{W_1(\mu_i,\mu_j)}{d(i,j)}.$7 hinges through deficit angles

$\kappa_{ij}=1-\frac{W_1(\mu_i,\mu_j)}{d(i,j)}.$8

and sectional curvature is $\kappa_{ij}=1-\frac{W_1(\mu_i,\mu_j)}{d(i,j)}.$9. The resulting simplicial Ricci tensor $\mu_t(i,j)$0 is attached to primal edges, while the discrete Ricci-flow equation is expressed through a volume-weighted average of fractional dual-edge length changes. This Regge–Ricci flow provides an algebraic, edge-based analogue of Hamilton’s equations in a hybrid primal–dual orthogonal basis (Miller et al., 2013).

A distinct 3-dimensional combinatorial formulation works directly with a PL metric $\mu_t(i,j)$1 on the edge set of a triangulated $\mu_t(i,j)$2-manifold. The combinatorial Ricci curvature at an edge is

$\mu_t(i,j)$3

and the total curvature functional is $\mu_t(i,j)$4. The unnormalized flow is

$\mu_t(i,j)$5

while the normalized flow is

$\mu_t(i,j)$6

A discrete Einstein metric is characterized by $\mu_t(i,j)$7. The theory proves that limits of the normalized flow are discrete Einstein metrics, and that convergence is exponential near such a metric under a spectral-gap condition $\mu_t(i,j)$8 (Ge et al., 2013).

Conboye and Miller construct scalar, sectional, and Ricci curvatures for piecewise-flat $\mu_t(i,j)$9-manifolds by integrating hinge-based curvature over dual volumes. For an edge $\frac{d}{dt}\mu_t(i,j)=-\kappa_{ij}(t)\mu_t(i,j),$0, the discrete Ricci curvature is

$\frac{d}{dt}\mu_t(i,j)=-\kappa_{ij}(t)\mu_t(i,j),$1

and the normalized piecewise-flat Ricci flow is

$\frac{d}{dt}\mu_t(i,j)=-\kappa_{ij}(t)\mu_t(i,j),$2

The paper reports convergence of these piecewise-flat curvatures, and of the induced Ricci flow, to their smooth counterparts on several test manifolds, including triangulations of the $\frac{d}{dt}\mu_t(i,j)=-\kappa_{ij}(t)\mu_t(i,j),$3-sphere, $\frac{d}{dt}\mu_t(i,j)=-\kappa_{ij}(t)\mu_t(i,j),$4, a Gowdy $\frac{d}{dt}\mu_t(i,j)=-\kappa_{ij}(t)\mu_t(i,j),$5-model, and a Nil$\frac{d}{dt}\mu_t(i,j)=-\kappa_{ij}(t)\mu_t(i,j),$6-model (Conboye et al., 2016).

Singularity handling enters through surgery. In the discrete Ricci flow with surgery for axially symmetric $\frac{d}{dt}\mu_t(i,j)=-\kappa_{ij}(t)\mu_t(i,j),$7-geometries, the edge-length system

$\frac{d}{dt}\mu_t(i,j)=-\kappa_{ij}(t)\mu_t(i,j),$8

is evolved until a Type-I neck pinch forms. The singular edge is removed, the manifold is capped by icosahedral fans, the two resulting lobes are remeshed by cubic-spline refinement, and each component is then evolved independently. The reported outcome is the expected Thurston decomposition into two $\frac{d}{dt}\mu_t(i,j)=-\kappa_{ij}(t)\mu_t(i,j),$9-type lobes, each subsequently collapsing toward a round $$w_{t+1}(u,v)=w_t(u,v)-\varepsilon\,\kappa_t(u,v)\,w_t(u,v),$$0-sphere geometry (Alsing et al., 2017).

5. Energy, stability, convergence, and prescribed curvature

Analytically, discrete Ricci flows are governed by several recurring structures: convex energy functionals, monotonicity formulas, maximum principles, and local Lipschitz dependence of curvature on the metric data. On surfaces in hyperbolic background geometry, the Ricci potential

$$w_{t+1}(u,v)=w_t(u,v)-\varepsilon\,\kappa_t(u,v)\,w_t(u,v),$$1

has positive-definite Hessian, is strictly convex, and decreases along the flow, yielding uniqueness of the zero-curvature metric and exponential convergence when such a metric exists (Ge et al., 2015). In the unified surface theory, the discrete Ricci energy plays the same structural role: its gradient is the curvature defect $$w_{t+1}(u,v)=w_t(u,v)-\varepsilon\,\kappa_t(u,v)\,w_t(u,v),$$2, and its Hessian is the curvature Jacobian (zhang et al., 2014).

Graph flows display analogous, but not identical, stability mechanisms. In the controlled network formulation, the total squared tracking error

$$w_{t+1}(u,v)=w_t(u,v)-\varepsilon\,\kappa_t(u,v)\,w_t(u,v),$$3

is shown to be nonincreasing under the closed-loop flow when $$w_{t+1}(u,v)=w_t(u,v)-\varepsilon\,\kappa_t(u,v)\,w_t(u,v),$$4, which gives Lyapunov stability. With imperfect target information, a composite Lyapunov functional including estimator and input errors remains nonincreasing under a modified estimator flow (Sandhu et al., 2019).

For Ollivier-type graph curvature more broadly, a significant development is the proof that discrete-time curvature is Lipschitz and piecewise regular in time for broad classes of local random walks, while continuous-time curvature admits a limit-free operator-theoretic formulation. These properties imply local existence and uniqueness for generalized continuous-time Ollivier–Ricci flows on finite graphs whenever the vector field is locally Lipschitz in the metric data and curvature (Fathi et al., 2022). On infinite graphs, prescribed-curvature flow further supplies global existence, uniqueness, and convergence on girth-$$w_{t+1}(u,v)=w_t(u,v)-\varepsilon\,\kappa_t(u,v)\,w_t(u,v),$$5 graphs under small-data or one-sided maximum-principle hypotheses (Hua et al., 8 Jun 2026).

Surgery-based weighted-graph flows offer a different route to convergence. In the piecewise-linear framework, curvature is frozen on each time interval and edges whose weight ratio exceeds a fixed threshold are deleted. If the chosen curvature is homogeneous, finitely many surgeries occur, after which each connected component reaches constant Ricci curvature (Ma et al., 21 May 2025). This contrasts with empirical graph Ricci-flow metrics used in alignment, where curvature uniformization is observed but a universal convergence theorem is explicitly absent (Ni et al., 2018).

These results delimit a central controversy in the field: convergence claims are highly model-dependent. Surface and prescribed-curvature formulations often support necessity-and-sufficiency theorems or energy-based proofs, whereas some network-oriented flows remain justified primarily by robustness experiments and heuristic analogy with smooth uniformization. The literature therefore supports both rigorous geometric analysis and pragmatic algorithmics, but not a single convergence doctrine.

6. Applications, limitations, and current directions

Applications of discrete Ricci curvature flow are unusually broad because the evolving metric can be interpreted either geometrically or algorithmically. In networks, controlled Ollivier flow is proposed as a mechanism to alter Boltzmann entropy and, through the fluctuation-theorem link described there, network functional robustness. The same study presents scale-free-network experiments in which injected input flips the direction of entropic change and tracks curvature–entropy co-evolution (Sandhu et al., 2019). In graph matching, the Ricci-flow metric supports landmark-based alignment of noisy graphs, including email, Internet, and protein interaction networks (Ni et al., 2018).

Community and core detection have become major graph-theoretic uses. Piecewise-linear Ricci curvature flows with surgeries are applied to community detection on Karate Club, College Football, Facebook, and LFR benchmarks, where the reported advantages are convergence of the iterative process and the absence of curvature recomputation at every iteration (Ma et al., 21 May 2025). Directed Ricci flow extends this logic to strongly connected subgraph extraction in weakly connected directed graphs, using artificial high-weight edges only as an augmentation device (Zhao et al., 5 Dec 2025). In information retrieval, Ricci-Filtration models a query and retrieved chunks as a graph, uses normalized discrete Ricci flow to separate noisy from structurally relevant chunks, and then passes the filtered set to a reranker (Qin et al., 13 Jun 2026).

For geometry processing, the surface literature emphasizes uniformization, conformal parameterization, texture mapping, surface registration, and geometric modeling. The hyperbolic surface flow is explicitly presented as an algorithmic method for finding circle-packing metrics of prescribed curvature, with potential applications in computer graphics and the study of discrete geometric structures on $$w_{t+1}(u,v)=w_t(u,v)-\varepsilon\,\kappa_t(u,v)\,w_t(u,v),$$6-manifolds via boundary circle packings (Ge et al., 2015). The unified surface framework is designed to improve flexibility, robustness, and implementation efficiency across mesh qualities and topologies (zhang et al., 2014). On polyhedral and CW-complex models, metric-, Haantjes-, and Forman-based Ricci flows have been linked to imaging, denoising, segmentation, and network intelligence (Saucan, 2019).

In higher dimensions, discrete Ricci flow with surgery is positioned as an explicit numerical realization of Thurston’s geometrization in a PL setting (Alsing et al., 2017). Piecewise-flat formulations are intended to converge to smooth Ricci flow under mesh refinement (Conboye et al., 2016), while discrete Einstein-metric flows provide a finite-dimensional analogue of Einstein-metric search on triangulated $$w_{t+1}(u,v)=w_t(u,v)-\varepsilon\,\kappa_t(u,v)\,w_t(u,v),$$7-manifolds (Ge et al., 2013). For discrete surfaces of revolution, normalized flow numerically approaches the explicit constant-Gaussian-curvature profiles supplied by integrable-systems constructions (Suda, 2023).

The present state of the subject is therefore best understood as stratified rather than unified. Graph flows are rapidly expanding into network science, machine learning, and retrieval systems; surface flows retain the strongest uniformization and energy theory; simplicial and piecewise-flat flows pursue faithful discretizations of continuum geometry and singularity handling. What all branches share is the operational principle that curvature drives metric deformation, but the meaning of “curvature,” “metric,” and “convergence” remains discretization-specific.