Discrete Ricci Flow: Methods and Applications

• Discrete Ricci Flow is a curvature-driven evolution of discretized metrics on geometric structures such as simplicial complexes, triangulated surfaces, weighted graphs, and Kähler manifolds.
• It employs various update rules—on edge lengths, conformal factors, or weight systems—to steer the geometry toward constant-curvature, Einstein, or structurally separated configurations.
• Practical applications span computational geometry and network science, enabling efficient simulation of geometric flows, handling of singularities, and convergence analysis in complex discrete settings.

Discrete Ricci flow denotes a family of curvature-driven evolutions in which a discretized metric is updated by a rule modeled on smooth Ricci flow or normalized Ricci flow. Across the literature, the evolving variable may be an edge-length assignment on a simplicial or piecewise-flat manifold, a conformal factor or circle-packing radius on a triangulated surface, a sequence of Kähler metrics produced by a backward-Euler-type iteration, or an edge-weight system on a weighted graph. The common objective is to encode Ricci-type curvature in a discrete state space and to evolve that state toward constant-curvature, Einstein, or structurally separated configurations, depending on the setting (Miller et al., 2013, zhang et al., 2014, Darvas et al., 2017, Bai et al., 2020).

1. Principal formulations

The term covers several non-equivalent constructions. In simplicial and Regge-calculus formulations, curvature is concentrated on codimension-$2$ simplices and the flow acts on primal or dual edge lengths. In surface theories, discrete curvature is usually an angle defect at vertices and the flow acts on discrete conformal variables. In variational and iterative approaches, the flow is replaced by a minimizing-movement scheme or a Ricci iteration. In graph-theoretic work, Ollivier-type or related curvatures drive updates of edge weights rather than a Riemannian metric tensor (Miller et al., 2013, zhang et al., 2014, Ma et al., 2012, Darvas et al., 2017, Bai et al., 2020).

Setting	Discrete metric variable	Representative evolution
Piecewise-flat / simplicial manifolds	edge lengths $\ell$	$$\left\langle \frac{1}{\lambda}\frac{\partial \lambda}{\partial t}\right\rangle_\ell=-Rc_\ell$$, $\dot l=-R$, $$\frac{1}{|\ell|}\frac{d|\ell|}{dt}=-Rc_\ell+\frac{1}{n}\widetilde R_S$$
Triangulated surfaces	conformal factors $u_i$, radii $r_i$, or $g_i=\sinh^2 r_i$	$\frac{du_i}{dt}=\bar K_i-K_i$, $\frac{dg_i}{dt}=-R_i g_i$
Variational / iterative analogues	minimizers or iterates $\ell$0	time-discrete minimization; $\ell$1
Weighted graphs	edge weights $\ell$2	$\ell$3, $\ell$4

This multiplicity is not accidental. Different discretizations preserve different aspects of the smooth theory: Regge-style constructions emphasize curvature concentration and dual volumes, circle-packing schemes emphasize discrete conformal geometry, minimizing-movement schemes emphasize variational structure, and graph formulations emphasize transport-based curvature and network geometry. A recurring misconception is that there is a single canonical discrete Ricci flow; the literature instead exhibits several inequivalent but mathematically precise analogues.

2. Piecewise-flat and simplicial manifold theories

A major strand treats a smooth manifold by a simplicial piecewise-flat approximation whose geometry is determined by a simplicial complex together with edge lengths and, in many formulations, a dual tessellation. In three dimensions the hinges are edges, and the fundamental curvature datum is the deficit angle

$\ell$5

which measures the failure of parallel transport around a small loop encircling the hinge to return a vector unchanged (Conboye et al., 2016).

In the Regge-calculus and discrete-exterior-calculus construction, a piecewise-flat simplicial lattice $\ell$6 is paired with its circumcentric dual lattice $\ell$7. Hybrid volumes formed from primal simplices and dual cells provide the support for discrete curvature. The hinge sectional curvature is

$\ell$8

and the Regge-Ricci flow is expressed as a simplicial equation

$\ell$9

obtained by averaging fractional rates of change of dual edges $$\left\langle \frac{1}{\lambda}\frac{\partial \lambda}{\partial t}\right\rangle_\ell=-Rc_\ell$$0 over the dual cell of a primal edge $$\left\langle \frac{1}{\lambda}\frac{\partial \lambda}{\partial t}\right\rangle_\ell=-Rc_\ell$$1 (Miller et al., 2013). This formulation is a direct discrete analogue of $$\left\langle \frac{1}{\lambda}\frac{\partial \lambda}{\partial t}\right\rangle_\ell=-Rc_\ell$$2, but written in terms of primal and dual edge geometry rather than tensor components.

A related piecewise-flat three-dimensional framework defines scalar, sectional, and Ricci curvatures directly on dual volumes. For a vertex dual volume $$\left\langle \frac{1}{\lambda}\frac{\partial \lambda}{\partial t}\right\rangle_\ell=-Rc_\ell$$3,

$$\left\langle \frac{1}{\lambda}\frac{\partial \lambda}{\partial t}\right\rangle_\ell=-Rc_\ell$$4

while the Ricci curvature along an edge $$\left\langle \frac{1}{\lambda}\frac{\partial \lambda}{\partial t}\right\rangle_\ell=-Rc_\ell$$5 is assembled from endpoint scalar curvature and sectional curvature orthogonal to $$\left\langle \frac{1}{\lambda}\frac{\partial \lambda}{\partial t}\right\rangle_\ell=-Rc_\ell$$6,

$$\left\langle \frac{1}{\lambda}\frac{\partial \lambda}{\partial t}\right\rangle_\ell=-Rc_\ell$$7

or, in a refined form, by weighting the endpoint contributions by the portions of $$\left\langle \frac{1}{\lambda}\frac{\partial \lambda}{\partial t}\right\rangle_\ell=-Rc_\ell$$8 lying in adjacent vertex volumes. The corresponding normalized piecewise-flat Ricci flow is

$$\left\langle \frac{1}{\lambda}\frac{\partial \lambda}{\partial t}\right\rangle_\ell=-Rc_\ell$$9

Computations on the $\dot l=-R$0-sphere, $\dot l=-R$1-cylinder, Gowdy manifold, and Nil-$\dot l=-R$2 geometry show convergence of the discrete curvatures and of the flow under refinement. An important negative result is that the single-hinge sectional curvature does not converge well in general, whereas the dual-volume-based multi-hinge curvature $\dot l=-R$3 does (Conboye et al., 2016).

Another three-dimensional PL theory defines edge curvature by

$\dot l=-R$4

regards a discrete Einstein metric as a PL metric satisfying

$\dot l=-R$5

and introduces both a second-order and a fourth-order curvature flow. The second-order flow

$\dot l=-R$6

is the direct analogue of smooth Ricci flow and its normalized version, while the fourth-order flow

$\dot l=-R$7

is the gradient flow of the quadratic curvature energy $\dot l=-R$8. Convergence of the normalized second-order flow to a nondegenerate limit implies a discrete Einstein metric, and convergence of the fourth-order flow implies the same under rank assumptions on $\dot l=-R$9 (Ge et al., 2013).

This strand also includes a diagonalized, edge-based discrete Ricci flow using a Forman-Ricci-type local construction. By associating a curvature quantity directly to each edge, the resulting system avoids large sparse matrix inversion at each time step and becomes numerically efficient on axisymmetric three-geometries (Alsing et al., 2017).

3. Surface, circle-packing, and variational formulations

On triangulated surfaces, discrete Ricci flow is usually formulated as a conformal evolution. For a triangular mesh $$\frac{1}{|\ell|}\frac{d|\ell|}{dt}=-Rc_\ell+\frac{1}{n}\widetilde R_S$$0, the discrete Gaussian curvature is the angle defect

$$\frac{1}{|\ell|}\frac{d|\ell|}{dt}=-Rc_\ell+\frac{1}{n}\widetilde R_S$$1

and the discrete conformal factor is written, depending on the background geometry, as

$$\frac{1}{|\ell|}\frac{d|\ell|}{dt}=-Rc_\ell+\frac{1}{n}\widetilde R_S$$2

The unified surface Ricci flow is

$$\frac{1}{|\ell|}\frac{d|\ell|}{dt}=-Rc_\ell+\frac{1}{n}\widetilde R_S$$3

where $$\frac{1}{|\ell|}\frac{d|\ell|}{dt}=-Rc_\ell+\frac{1}{n}\widetilde R_S$$4 is the prescribed target curvature. The associated Ricci energy

$$\frac{1}{|\ell|}\frac{d|\ell|}{dt}=-Rc_\ell+\frac{1}{n}\widetilde R_S$$5

has gradient equal to the curvature error, and the symmetry $$\frac{1}{|\ell|}\frac{d|\ell|}{dt}=-Rc_\ell+\frac{1}{n}\widetilde R_S$$6 makes the defining $$\frac{1}{|\ell|}\frac{d|\ell|}{dt}=-Rc_\ell+\frac{1}{n}\widetilde R_S$$7-form closed. The same framework unifies tangential circle packing, Thurston circle packing, inversive distance circle packing, discrete Yamabe flow, virtual radius circle packing, and mixed-type schemes in Euclidean, hyperbolic, and spherical background geometries (zhang et al., 2014).

For hyperbolic background geometry, a circle packing metric is a positive radius function $$\frac{1}{|\ell|}\frac{d|\ell|}{dt}=-Rc_\ell+\frac{1}{n}\widetilde R_S$$8, with edge lengths determined by

$$\frac{1}{|\ell|}\frac{d|\ell|}{dt}=-Rc_\ell+\frac{1}{n}\widetilde R_S$$9

A modified combinatorial Gaussian curvature is defined by

$u_i$0

and the Ricci flow becomes

$u_i$1

or, in logarithmic coordinates $u_i$2,

$u_i$3

The principal convergence theorem states that the flow converges if and only if there exists a circle packing metric with zero curvature. A stronger theorem proves exponential convergence if the initial modified curvatures satisfy $u_i$4, without assuming the existence of a zero-curvature metric or invoking Thurston’s combinatorial-topological condition as a hypothesis (Ge et al., 2015).

A different surface formulation replaces an explicit flow equation by a time-discretized variational scheme. On the “football” surface

$u_i$5

with scalar curvature $u_i$6 and total area $u_i$7, the evolving metric is written $u_i$8. With time step $u_i$9, one constructs $r_i$0 as a minimizer of

$r_i$1

uses a Chen–Li Moser-type inequality to bound the exponential term, and passes to a weak limit. Under the symmetry assumption $r_i$2, this yields existence of a weak approximated Ricci flow on any finite time interval. A modified flow,

$r_i$3

admits an analogous existence theorem without the symmetry hypothesis (Ma et al., 2012).

Discrete Ricci flow also appears on discrete surfaces of revolution as an intrinsic evolution of the discrete first fundamental form. If

$r_i$4

then the unnormalized flow is

$r_i$5

and the normalized flow replaces $r_i$6 by $r_i$7, where

$r_i$8

The weighted area is preserved, and numerical examples show convergence toward explicit discrete constant-Gaussian-curvature surfaces of revolution, including sphere-type, spindle-type, pseudosphere-type, cone, and cusp limits. The paper is explicit that it does not prove a general global convergence theorem analogous to the smooth one (Suda, 2023).

4. Ricci iteration and Kähler discrete time

In Kähler geometry, a discrete Ricci flow often takes the form of a Ricci iteration rather than an edge-length or conformal-factor evolution. The classical iteration is

$r_i$9

and in the Kähler setting, for $g_i=\sinh^2 r_i$0, the time-$g_i=\sinh^2 r_i$1 Ricci iteration is defined by

$g_i=\sinh^2 r_i$2

When $g_i=\sinh^2 r_i$3, this is exactly the original Ricci iteration, and it is explicitly interpreted as a backward-Euler discretization of the normalized Kähler-Ricci flow (Darvas et al., 2017).

The principal convergence theorem states that if $g_i=\sinh^2 r_i$4 admits a Kähler-Einstein metric and the Kähler class equals $g_i=\sinh^2 r_i$5, then there exist holomorphic diffeomorphisms $g_i=\sinh^2 r_i$6 such that

$g_i=\sinh^2 r_i$7

smoothly. The result holds for every fixed $g_i=\sinh^2 r_i$8, so the discrete scheme inherits the same asymptotic behavior as the continuous Kähler-Ricci flow in the Kähler-Einstein setting (Darvas et al., 2017).

The proof is variational. The Ding functional and Mabuchi $g_i=\sinh^2 r_i$9-energy are minimized by Kähler-Einstein metrics and decrease along the iteration, while compactness is obtained in the $\frac{du_i}{dt}=\bar K_i-K_i$0-metric completion $\frac{du_i}{dt}=\bar K_i-K_i$1 of the space of Kähler potentials. The role of holomorphic automorphisms is essential: when nontrivial holomorphic vector fields are present, convergence holds only modulo $\frac{du_i}{dt}=\bar K_i-K_i$2. The sphere case yields a discrete proof of uniformization of $\frac{du_i}{dt}=\bar K_i-K_i$3, with convergence to the round metric after composing with Möbius transformations. This formulation is discrete in time but not combinatorial; its state space remains infinite-dimensional and smooth, yet it is an exact discrete analogue of Ricci flow in the sense of time stepping.

5. Graph-theoretic Ricci flows

On weighted graphs, discrete Ricci flow is typically driven by Ollivier-type curvature. For an undirected weighted graph, local probability measures $\frac{du_i}{dt}=\bar K_i-K_i$4 or $\frac{du_i}{dt}=\bar K_i-K_i$5 are defined at each vertex, the graph distance $\frac{du_i}{dt}=\bar K_i-K_i$6 is the shortest-path metric induced by edge weights, and the curvature along an edge $\frac{du_i}{dt}=\bar K_i-K_i$7 is

$\frac{du_i}{dt}=\bar K_i-K_i$8

or, in the Ollivier–Lin–Lu–Yau limit form,

$\frac{du_i}{dt}=\bar K_i-K_i$9

A widely used discrete-time update is

$\frac{dg_i}{dt}=-R_i g_i$0

often followed by normalization of total edge-weight mass, while a continuous-time normalized flow is

$\frac{dg_i}{dt}=-R_i g_i$1

This flow preserves $\frac{dg_i}{dt}=-R_i g_i$2 when the initial total weight is normalized. Under the paper’s exit condition that each edge remains the shortest path connecting its endpoints over time, existence and uniqueness hold globally. On finite star graphs with at least three leaves and $\frac{dg_i}{dt}=-R_i g_i$3, the normalized flow converges to the constant-weighted star; on finite paths it converges, after contractions, to a path of length $\frac{dg_i}{dt}=-R_i g_i$4 (Bai et al., 2020).

The curvature theory underlying such flows has itself been developed abstractly. For locally finite weighted graphs with general random walks, discrete-time Ollivier-Ricci curvature is piecewise regular in the walk parameter, piecewise affine for time-affine walks, concave for time-affine walks, and Lipschitz continuous. Continuous-time curvature admits a limit-free formulation and supports a generalized continuous-time Ollivier-Ricci curvature flow

$\frac{dg_i}{dt}=-R_i g_i$5

for which unique short-time solutions follow from Picard–Lindelöf once the vector field is locally Lipschitz. This framework is foundational: it establishes well-posedness for a broad class of graph curvature flows rather than privileging a single discrete dynamic (Fathi et al., 2022).

Directed graphs require an asymmetric transport geometry. For a weighted directed graph $\frac{dg_i}{dt}=-R_i g_i$6, the outward $\frac{dg_i}{dt}=-R_i g_i$7-lazy one-step random walk is

$\frac{dg_i}{dt}=-R_i g_i$8

the directed distance $\frac{dg_i}{dt}=-R_i g_i$9 is the shortest-path length over directed paths, and the curvature on a directed edge $\ell$00 is

$\ell$01

The continuous flow

$\ell$02

has a unique global solution for strongly connected directed graphs, and the discrete flow

$\ell$03

is uniquely solvable for $\ell$04. For weakly connected directed graphs, the method adds a minimal set of artificial edges with very large weights to enforce strong connectivity, then removes those edges during the final surgery stage (Zhao et al., 5 Dec 2025).

6. Surgery, singularities, and applications

One line of work uses discrete Ricci flow to model singular behavior and continuation past singular times. In an axially symmetric simplicial neckpinch geometry, a diagonalized edge-based discrete Ricci flow reproduces finite-time Type-$\ell$05 neck pinch behavior. At approximately $\ell$06, the surgery procedure removes the pinched axial edge $\ell$07, caps each open end with an icosahedron, remeshes each lobe by cubic spline interpolation, and resumes the flow on the disconnected components. The two lobes then evolve toward spherical end states, giving an explicit numerical realization of Thurston’s geometrization procedure in a PL setting (Alsing et al., 2017).

A related but more synthetic theory defines super Ricci flow for time-dependent finite weighted graphs. The graph structure may change at finitely many singular times by collapse, spawning, or edge deletion and creation. The flow is characterized equivalently by a dynamic Bochner inequality,

$\ell$08

a gradient estimate for the heat flow, a transport contraction estimate for the dual heat flow, and dynamic convexity of entropy along discrete optimal transport paths. The framework is built precisely so that the flow can continue through graph singularities and remains consistent with classical super Ricci flows in a discrete-to-continuum limit (Erbar et al., 2018).

Applications in network science rely on the same shrink-positive / stretch-negative mechanism. In community detection, discrete Ricci flow reweights edges until inter-community bridges become heavy and intra-community edges become light; a subsequent thresholding step, called network surgery, removes high-weight edges and reveals communities. On stochastic block model, LFR, and emergent geometrical network benchmarks, as well as on Karate club, football, political books, political blogs, Facebook ego, and Email-EU-core, the method is reported to be effective; Sinkhorn transport gives similar clustering quality and is about four times faster than exact transport in the reported tests (Ni et al., 2019). In network alignment, the shortest-path metric after Ricci flow is used as a robust landmark-distance signature; RF-ATD is reported to be up to $\ell$09 faster than RF-OTD, and many experiments report over $\ell$10 matching accuracy (Ni et al., 2018). In entropy control, discrete Ricci flow is embedded in a feedback law

$\ell$11

and a Lyapunov analysis is given for Ollivier-Ricci-based discretization (Sandhu et al., 2019).

Recent work extends the same paradigm to directed core extraction, retrieval-augmented generation, and learned feature geometry. For directed graphs, a flow-and-surgery pipeline detects strongly connected core subgraphs and is reported to outperform baseline centrality methods on at least two of three structural metrics on Physicians, Elegans, and Human protein networks (Zhao et al., 5 Dec 2025). In Ricci-Filtration for RAG reranking, a query–chunk graph is evolved by

$\ell$12

and chunks whose query-edge weights exceed a surgery threshold are discarded before reranking; the method improves several QA benchmarks, although it is slower than a plain cross-encoder and is less effective on HotpotQA and some MuSiQue settings (Qin et al., 13 Jun 2026). In neural representation learning, graph approximations of feature manifolds are tracked layer by layer, and the observed evolution is compared with discrete Ricci flow through local Ricci evolution coefficients; experiments on more than $\ell$13 feedforward networks suggest that feature geometry evolves in a Ricci-flow-like manner, with class separability coinciding with emergence of community structure (Hehl et al., 26 Sep 2025).

This range of constructions suggests a broad unifying pattern: discrete Ricci flow is best understood not as one algorithm, but as a class of curvature-driven discrete evolutions whose specific meaning depends on the underlying discretization of metric, curvature, and transport. In every major formulation, however, the decisive ingredients are the same—an explicit curvature notion, a metric update rule, and a compactness, convergence, or structural-separation mechanism connecting the discrete dynamics to a geometric target.