Forman Curvature in Discrete Networks

• Forman curvature is a discrete, edge-based analogue of Ricci curvature derived from combinatorial methods, quantifying local geometric properties in networks.
• It extends to multiple network types—including weighted, directed, temporal, and higher-order structures—offering a versatile and computationally efficient tool for network analysis.
• Augmented variants incorporate cycle and higher-dimensional cell information, enhancing sensitivity to local redundancy, community structure, and overall network robustness.

Forman curvature, often called Forman–Ricci curvature in network science, is a combinatorial, edge-based analogue of Ricci curvature derived from Forman’s discretization of the Bochner–Weitzenböck formula. It was introduced on weighted cell complexes and later adapted to undirected, directed, temporal, multilayer, and higher-order network settings, where it functions as a local but geometrically meaningful measure of how an edge or cell sits inside surrounding combinatorial structure. In graphs it is fundamentally an edge quantity; node, face, and higher-dimensional versions are obtained either by aggregation or by working directly on CW or simplicial complexes. Across the literature, Forman curvature is valued for combining geometric interpretation with computational simplicity, while subsequent work has clarified where unaugmented, augmented, weighted, and higher-order variants coincide and where they diverge (Sreejith et al., 2016, Sreejith et al., 2016, Souza et al., 30 Apr 2025, Yamada, 15 May 2026).

1. Geometric origin and combinatorial foundations

Forman curvature enters discrete geometry through a Bochner–Weitzenböck-type decomposition. In the weighted cell-complex setting, the combinatorial Riemann–Laplace operator satisfies

$\Box_p = B_p + F_p,$

where $B_p$ is the combinatorial Bochner Laplacian and $F_p$ is a diagonal curvature operator; for a $p$-cell $\alpha^p$, the curvature function is $$\mathcal{F}_p(\alpha)=\langle F_p(\alpha),\alpha\rangle$$. The construction depends on three incidence classes: cofaces of dimension $p+1$, faces of dimension $p-1$, and same-dimensional parallel cells. In this framework, two $p$-cells are parallel if they share a common $(p+1)$-coface or a common $B_p$0-face, but not both [(Weber et al., 2016); (Bloch, 2014)].

On abstract simplicial complexes, the combinatorial content of the definition becomes especially transparent. For a $B_p$1-cell $B_p$2, with $B_p$3 the set of $B_p$4-cells containing $B_p$5 and $B_p$6 the parallel neighbors of $B_p$7, the high-order Forman curvature is

$B_p$8

A set-theoretic reformulation expresses the same quantity entirely through vertex-neighborhood intersections:

$B_p$9

This formula is central in later algorithmic work because it avoids explicit construction of high-order simplex adjacency and makes curvature locally computable cell by cell (Souza et al., 2023).

A distinct 2-dimensional development arises in ranked posets and polyhedral surfaces. Because Forman’s original edge-only combinatorial Ricci curvature does not satisfy a Gauss–Bonnet analog in dimension $F_p$0, Bloch introduced curvature functions on vertices, edges, and faces, obtaining the identity

$F_p$1

This redistributes curvature across ranks while preserving the incidence-theoretic spirit of the original construction. A plausible implication is that, in low-dimensional combinatorial geometry, Forman curvature is best understood not only as an edge observable but also as part of a broader curvature bookkeeping scheme (Bloch, 2014).

2. Graph formulations and node-level extensions

In simple unweighted graphs, the edge formula collapses to the familiar degree expression

$F_p$2

for an edge $F_p$3. This makes plain Forman curvature depend only on the degrees of the two endpoints. In weighted graphs, the full formula directly incorporates node weights, edge weights, and incident edges at the endpoints, so the curvature remains local while accommodating weighted data (Sreejith et al., 2016, Fesser et al., 2023).

The structural interpretation in network science is tied to branching at the endpoints of an edge. When an edge is surrounded by many adjacent edges at its endpoints, the negative contributions dominate and the curvature becomes more negative. In this sense, Forman curvature captures the extent of spreading at the ends of an edge and reflects a local divergence or branching structure. The same sign convention underlies most empirical work: highly negative curvature typically marks edges embedded in strongly spreading neighborhoods, whereas less negative or positive values correspond to locally constrained configurations (Sreejith et al., 2016).

Because curvature is fundamentally edge-based, node-level quantities are introduced by aggregation. Two natural definitions are

$F_p$4

and

$F_p$5

The first is the normalized average over incident edges; the second is an unnormalized accumulated curvature. In unweighted networks the literature considers both combinatorial node weights $F_p$6 and degree weights $F_p$7, but systematic evaluation found that the unnormalized node curvature with combinatorial node weights is the more effective practical choice, giving the strongest negative correlation with node degree and node betweenness and the best performance in node-removal robustness experiments (Sreejith et al., 2016).

This nonuniqueness of the node extension is one of the central interpretive cautions surrounding Forman curvature. The underlying edge quantity is canonical within the chosen combinatorial model, but scalar-like node curvature is an aggregation choice rather than a uniquely determined object. That distinction remains important in later work on directed, multiplex, and higher-order networks (Sreejith et al., 2016).

3. Cycle augmentation and higher-order sensitivity

A recurring criticism of plain graph Forman curvature is that, in its simplest unweighted form, it ignores local cycle structure. This limitation is explicit in work on community detection, which writes the un-augmented edge curvature as

$F_p$8

and observes that this degree-only quantity does not distinguish whether an edge lies inside a densely interconnected region or acts as a bridge when endpoint degrees are similar. Augmented Forman–Ricci curvature (AFRC) addresses this by “filling in” cycles with 2-dimensional faces and applying Forman’s cell-complex formula to the resulting augmented complex (Fesser et al., 2023).

The most important special cases are short-cycle augmentations. For triangles,

$F_p$9

so edges contained in many triangles receive a curvature boost. In GNN-oriented notation, the same principle appears as

$p$0

and with quadrangles

$p$1

These formulas encode the idea that short cycles capture local redundancy of paths: triangles are particularly informative in unipartite graphs, whereas four-cycles become important in bipartite or nearly bipartite settings (Fesser et al., 2023, Fesser et al., 2023).

Empirically, $p$2 frequently behaves similarly to Ollivier–Ricci curvature in triangle-rich community-structured graphs, while $p$3 can become more appropriate in bipartite settings. AFRC-based community detection was reported to be competitive with an ORC-based approach and “orders of magnitude faster” in the cited experiments. At the same time, the augmentation literature is not uniform. In persistent-homology filtrations on non-quasiconvex networks, triangle augmentation was recommended, whereas plain curvature was found to omit too much information and quasiconvex pentagon-augmented curvature was judged too rough an approximation, producing significant distortion (Fesser et al., 2023, Iváñez, 2022).

A further complication is that “augmented” formulas are exact only under specific structural assumptions. On temporal prism complexes, the original CW-complex curvature and the augmented network-science variant coincide under uniform weights,

$p$4

but diverge generically when temporal edges carry interval-dependent weights. This suggests that cycle augmentation is not a single universal refinement; it is a family of approximations whose fidelity depends on both combinatorial regime and weighting scheme (Yamada, 15 May 2026).

4. Directed, hypergraph, multiplex, temporal, and mosaic generalizations

The extension to directed graphs preserves the edge-centric character of the construction while enforcing orientation compatibility. For a directed edge $p$5, one counts only incoming edges to $p$6 and outgoing edges from $p$7, leading to node aggregates

$p$8

and a net flow quantity

$p$9

This directional version is designed to reflect local flow through the edge and has been studied on model, real, weighted, and spatial directed networks (Saucan et al., 2018).

Hypergraph generalizations keep the emphasis on edge-like relations rather than converting all higher-order interactions to ordinary graph edges. In undirected unweighted hypergraphs,

$\alpha^p$0

so curvature quantifies the trade-off between hyperedge size and the degree of participation of hyperedge vertices in other hyperedges. Directed hypergraphs split this into forward and reverse components, for example

$\alpha^p$1

making it possible to distinguish flow-through structure from redundancy or replaceability. In related work on polyhedral hypernetworks, Forman curvature is used on weighted simplicial or polyhedral complexes and is paired with a structure-preserving embedding of a hypernetwork in $\alpha^p$2 (Leal et al., 2018, Saucan et al., 2018).

For multiplex graphs, the construction is lifted to layer-specific state vertices $\alpha^p$3 and both intra-layer and inter-layer edges. The paper on doubly weighted multiplex graphs defines curvature on the state graph and uses the inter-layer aggregate

$\alpha^p$4

as a vertex evaluation across layers. The stated interpretation is that the gap between $\alpha^p$5 and its uniform-layer baseline measures how differently the same vertex behaves across layers (Yamada, 24 Apr 2025).

Temporal networks motivate a different higher-dimensional model. A contact sequence is lifted to a spatiotemporal prism complex whose 1-simplices fall into three classes: spatial edges, temporal edges, and diagonal edges. On this complex, the original weighted CW-complex Forman curvature and the augmented variant agree exactly for spatial edges under the default spatial weights, but they differ on temporal and diagonal edges whenever the temporal weight function satisfies $\alpha^p$6. The cited experiments reported disagreement on $\alpha^p$7–$\alpha^p$8 of the $\alpha^p$9-simplices, predominantly the temporal and diagonal ones, while retaining strong Pearson correlation $$\mathcal{F}_p(\alpha)=\langle F_p(\alpha),\alpha\rangle$$0 (Yamada, 15 May 2026).

Specialized geometric domains have also produced modified variants. For irregular convex mosaics, the classical edge formula

$$\mathcal{F}_p(\alpha)=\langle F_p(\alpha),\alpha\rangle$$1

is refined by splitting degree into regular and irregular parts,

$$\mathcal{F}_p(\alpha)=\langle F_p(\alpha),\alpha\rangle$$2

so that curvature distributions become sensitive to straight versus skewed crack geometry in fracture patterns (Gupta et al., 28 May 2026).

5. Computation, analytic characterizations, and global constraints

A major reason for the visibility of Forman curvature in applied work is computational tractability. In clique complexes and abstract simplicial complexes, the set-theoretic formula

$$\mathcal{F}_p(\alpha)=\langle F_p(\alpha),\alpha\rangle$$3

reduces curvature evaluation to local neighborhood intersections. FastForman was introduced around this observation, with the explicit claim that the reformulation reveals previous computational bottlenecks and improves both time and memory usage relative to earlier implementations (Souza et al., 2023).

On Vietoris–Rips filtrations, the same local viewpoint supports dynamic updates rather than repeated global recomputation. The decomposition

$$\mathcal{F}_p(\alpha)=\langle F_p(\alpha),\alpha\rangle$$4

leads to incremental rules in which, when a simplex is inserted, nearby curvatures change only by $$\mathcal{F}_p(\alpha)=\langle F_p(\alpha),\alpha\rangle$$5 or $$\mathcal{F}_p(\alpha)=\langle F_p(\alpha),\alpha\rangle$$6. This underlies the “data geometrization” workflow, where point clouds or tabular data are transformed into curvature profiles as a function of the cutoff distance and only affected local neighborhoods are updated across the filtration (Souza et al., 30 Apr 2025).

Analytically, one line of research characterizes Forman curvature lower bounds through the Hodge Laplacian semigroup. In that framework,

$$\mathcal{F}_p(\alpha)=\langle F_p(\alpha),\alpha\rangle$$7

The same work proves that, on edges, Ollivier curvature equals the maximum Forman curvature obtained by varying admissible $$\mathcal{F}_p(\alpha)=\langle F_p(\alpha),\alpha\rangle$$8-cells,

$$\mathcal{F}_p(\alpha)=\langle F_p(\alpha),\alpha\rangle$$9

and explicitly warns that its Forman curvature notion does not coincide with Forman’s original weighted definition, although the two agree when $p+1$0. This establishes a precise bridge between optimal transport and higher-cell augmentation (Jost et al., 2021).

Positive curvature imposes strong global restrictions in some combinatorial surface settings. For tessellations on surfaces, the unit-weight edge curvature simplifies to

$p+1$1

Under the condition $p+1$2 for every edge, the tessellation must be finite; moreover, the cited classification gives exactly $p+1$3 such graphs up to isomorphism, all planar. This is a discrete Bonnet–Myers-type phenomenon in a particularly rigid combinatorial regime (2002.03550).

6. Empirical behavior, applications, and interpretation

In model and real undirected networks, most nodes and edges were found to have negative Forman curvature. The distributions are narrow in Erdős–Rényi and Watts–Strogatz networks and broad in Barabási–Albert, power-law-cluster, and many real networks. Forman curvature is significantly negatively correlated with degree and several centrality measures, but is generally uncorrelated with clustering coefficient. In robustness experiments, targeted deletion of nodes or edges with highly negative curvature fragments networks faster than random removal, although removal by degree or betweenness is often still more disruptive (Sreejith et al., 2016).

A more refined edge-level comparison showed that Forman curvature is a better indicator of edge importance than embeddedness or dispersion, especially for large-scale connectivity, while usually remaining inferior to edge betweenness except in networks with explicit geometric structure such as NGF networks. For node-level use, the unnormalized definition with combinatorial node weights $p+1$4 was recommended. A central interpretive caution from this literature is that negative curvature is not “bad”: highly negative edges and nodes are often precisely the structurally important ones whose removal disrupts connectivity most strongly (Sreejith et al., 2016).

In directed networks, analogous patterns persist. Most edges are negatively curved, curvature distributions distinguish random from scale-free organization, and directed edge curvature is negatively correlated with edge betweenness. Node-level in-curvature and out-curvature are strongly negatively correlated with in-degree, out-degree, betweenness, and often PageRank; curvature-based targeted removal can reduce communication efficiency rapidly, sometimes performing comparably to or better than edge-betweenness attacks in weighted real networks such as US Airport and $p+1$5 elegans (Sreejith et al., 2016).

Applications have expanded well beyond classical network diagnostics. In community detection, AFRC frequently provides sufficient structural signal to serve as a computationally cheaper alternative to ORC-based methods, with $p+1$6 effective for unipartite stochastic block models and $p+1$7 valuable in bipartite or hierarchical bipartite settings (Fesser et al., 2023). In graph learning, AFRC is used as a curvature signal for both over-smoothing and over-squashing: high curvature identifies edges whose endpoint neighborhoods overlap strongly, while low curvature marks bottlenecks. The AFR-$p+1$8 rewiring schemes add edges around low-curvature edges and remove high-curvature ones, with complexity

$p+1$9

and the reported experiments show strong empirical performance with substantially reduced cost (Fesser et al., 2023).

Higher-order and data-analytic uses are equally varied. On Vietoris–Rips complexes, curvature profiles were used to distinguish random geometric graph classes, to separate Datasaurus randomizations that preserve basic summary statistics but change geometry, and to reveal diagnostically useful cutoff intervals in breast cancer datasets (Souza et al., 30 Apr 2025). In collider data, Forman curvature distributions on weighted Vietoris–Rips complexes of reconstructed leptons responded to MET cuts, filtration scale, and luminosity, and served as effective local topological features for discriminating BSM benchmarks from SM background (Beuria, 2023).

Taken together, these results support a consistent interpretation. Plain Forman curvature is fast, local, and geometrically suggestive, but it can be too coarse when cycle structure, higher-order faces, or nonuniform weights are central. Augmented, directed, temporal, multiplex, and hypergraph variants enlarge its expressive range, yet they also make explicit that there is no single universally preferred formulation. The stable core of the subject is the edge- or cell-based measurement of local combinatorial geometry; the active frontier concerns which augmentation, weighting, or higher-order lift best matches the structure under study (Iváñez, 2022, Yamada, 15 May 2026, Yamada, 24 Apr 2025).