Graph Curvature: Combinatorial & Metric Methods

• Graph curvature is a measure quantifying how discrete structures depart from flatness using combinatorial definitions and metric (Bakry–Émery) approaches.
• Combinatorial methods use vertex and face degree formulas to derive Gauss–Bonnet analogues, while metric techniques impose curvature-dimension conditions to control topological invariants.
• These insights inform network analysis by linking local curvature bounds to global properties like cycle homology, finite fundamental groups, and diameter limitations.

Graph curvature quantifies to what degree a discrete structure such as a graph deviates from “flatness” or exhibits properties analogous to smooth Riemannian manifolds with positive, zero, or negative curvature. Both combinatorial and metric approaches to graph curvature have yielded deep results, rigorous analogies to classical theorems from differential geometry, and connections to topology, group theory, and network analysis.

1. Combinatorial Curvature: Classical and Path-Based Formulations

The foundational notion of combinatorial curvature is defined for planar graphs or tessellations. At a vertex $v$ of degree $d_v$ incident to faces $f$ of degree $d_f$, the discrete (angle-defect) curvature is given by

$$\kappa(v) = 1 - \frac{d_v}{2} + \sum_{f\sim v} \frac{1}{d_f},$$

where the sum is over all faces containing $v$ (Kamtue, 2018). This formula reflects the angular deficit at $v$ relative to the Euclidean case, paralleling the smooth Gauss–Bonnet notion. For planar graphs, a Gauss–Bonnet-type theorem holds:

$\sum_{v\in V} \kappa(v) = 2,$

when the graph is drawn on the sphere (Keller, 2011).

Combinatorial curvature controls global properties:

• Non-positive curvature ($\kappa(v) \le 0$) at all vertices implies the graph is infinite and “locally” resembles a tessellation of the plane (Keller, 2011).
• Positive curvature everywhere restricts the graph to be finite (discrete Bonnet–Myers) and tightly controls its topology.

Path homology provides a combinatorial homological framework for graphs. The first path homology group $H_1(G;F)$ is isomorphic to the usual cycle space $d_v$0 modulo the subspace generated by all triangles and squares:

$d_v$1

(Kempton et al., 2017). Thus, cycles of length at least 5 give rise to nontrivial homology not generated by small cycles.

2. Metric and Bakry–Émery Curvature Dimension

The Bakry–Émery curvature dimension condition ($d_v$2) is a pivotal metric-analytic structure for graphs. For function $d_v$3, the Laplacian and carré du champ operators are defined as:

$d_v$4

The $d_v$5 condition at $d_v$6 is

$d_v$7

A graph with $d_v$8 everywhere is said to have a positive lower curvature bound.

Key results:

• Bochner-Type Theorem: In a finite graph, $d_v$9 with $f$0 implies $f$1 for any field $f$2 of characteristic zero. Thus positive curvature annihilates the first (path) homology group (Kempton et al., 2017).
• Myers-Type Theorem: $f$3 with $f$4 implies the combinatorial fundamental group $f$5 (with triangles and squares as null-homotopies) is finite, precluding infinite covers that “preserve” all 3- and 4-cycles (Kempton et al., 2017).
• Diameter Bound: Positive curvature implies a Bonnet–Myers-type upper bound on the diameter; explicit bounds depend on the setting (combinatorial, resistance-based, or equilibrium-measure curvature) (Kamtue, 2018, Steinerberger, 2022, Devriendt et al., 2023).

3. Coverings, Gain Graphs, and Algebraic-Topological Interpretation

Gain graphs associate to each oriented edge a group element, tracking how cycles “wind” in the group. For an abelian gain graph with gain function $f$6 (with $f$7 abelian),

$f$8

a cycle $f$9 is balanced if $d_f$0. Covered graphs $d_f$1 record precisely which cycles “lift” to closed walks in the cover.

The fundamental group $d_f$2 is presented by oriented edges subject to relations given by spanning tree edges and all triangle and square cycles being trivial. Its abelianization is isomorphic to the (first) path homology group:

$d_f$3

(Kempton et al., 2017).

If $d_f$4 satisfies $d_f$5 with $d_f$6, there can be no nontrivial infinite covering preserving all 3- and 4-cycles (no infinite abelian gain graph with all small cycles balanced but some large cycles unbalanced), yielding homological and homotopical rigidity.

4. Comparison of Combinatorial and Metric Approaches

The metric (Bakry–Émery, Ollivier–Ricci, equilibrium-measure, resistance) and combinatorial (cycle space, gain graph, cycle quotient) approaches interlock at several levels:

• Vanishing & Finiteness: Positive Bakry–Émery curvature ($d_f$7, $d_f$8) forces trivial $d_f$9 (vanishing first homology) and finite fundamental group, paralleling Bochner and Myers results in Riemannian geometry (Kempton et al., 2017).
• Obstructions: Non-positive curvature (in either sense) is compatible with infinite covers, nontrivial first homology, and “large” topology.
• Cycle Space Quotients: The combinatorial approach interprets $$\kappa(v) = 1 - \frac{d_v}{2} + \sum_{f\sim v} \frac{1}{d_f},$$0 as $$\kappa(v) = 1 - \frac{d_v}{2} + \sum_{f\sim v} \frac{1}{d_f},$$1triangles, squares$$\kappa(v) = 1 - \frac{d_v}{2} + \sum_{f\sim v} \frac{1}{d_f},$$2, clarifying precisely which homological features are “killed” by small cycles—exactly those detected by the metric vanishing theorems.
• Duality: The path homology/curvature dichotomy matches Bochner's formula: positive curvature kills 1-forms, so first Betti vanishes (Kempton et al., 2017).

5. Extremal and Illustrative Cases

• Trees: Have trivial path homology $$\kappa(v) = 1 - \frac{d_v}{2} + \sum_{f\sim v} \frac{1}{d_f},$$3 and trivial $$\kappa(v) = 1 - \frac{d_v}{2} + \sum_{f\sim v} \frac{1}{d_f},$$4 (combinatorial fundamental group), although they may not have $$\kappa(v) = 1 - \frac{d_v}{2} + \sum_{f\sim v} \frac{1}{d_f},$$5 with $$\kappa(v) = 1 - \frac{d_v}{2} + \sum_{f\sim v} \frac{1}{d_f},$$6 globally (the failure of positive curvature at degree-$$\kappa(v) = 1 - \frac{d_v}{2} + \sum_{f\sim v} \frac{1}{d_f},$$7 leaves).
• 5-cycle with diagonals: Can exhibit nonnegative curvature at every vertex, but still $$\kappa(v) = 1 - \frac{d_v}{2} + \sum_{f\sim v} \frac{1}{d_f},$$8; this demonstrates the necessity of $$\kappa(v) = 1 - \frac{d_v}{2} + \sum_{f\sim v} \frac{1}{d_f},$$9 (strict positivity) rather than mere nonnegativity for the vanishing theorem (Kempton et al., 2017).
• Cayley graphs of compact groups: Often satisfy $v$0 and thus have small $v$1 and finite $v$2, leading to strong expansion and small diameter properties.
• Regular graphs without $v$3 or $v$4 as subgraphs: The sign of Bakry–Émery curvature and Ollivier–Ricci curvature agrees and is determined by a local combinatorial invariant (the “non-linking number” $v$5) (Ralli, 2017).
• Metric Graphs: For continuum (metric) graphs, weak Bakry–Émery-type conditions (gradient estimates and evolution variational inequalities) provide a direct analog of curvature dimension but with quantifiable “loss” arising from vertex branching (Krautz, 17 Dec 2025).

6. Synthesis of the Combinatorial–Metric Interaction

The interplay between curvature and algebraic/topological invariants of a graph is formalized as follows (Kempton et al., 2017):

• The metric/Gaussian–Bochner approach, via discrete $v$6 and curvature-dimension inequalities, imposes combinatorial restrictions on cycle generators and coverings, resulting in algebraic consequences (vanishing of first homology and finiteness of $v$7).
• The combinatorial approach identifies $v$8 with the cycle space modulo the span of all small cycles, implements coverings via gain graphs, and realizes $v$9 as a canonical group presentation.
• These perspectives dovetail: positive curvature conditions force strong topological restrictions (e.g., all non-contractible cycles are “killed” modulo small cycles), while combinatorial and path-homological constructions make precise the obstructions due to negative or zero curvature.

A plausible implication is that further refinements of the curvature-dimension framework (e.g., Bakry–Émery with finite dimension parameter, metric measure generalizations, or higher path homology) will sharpen the classification of graph topology, spectral invariants, and covering spaces, as suggested by the continuum-dimension case and the interplay of gain graphs and cycle spaces (Kempton et al., 2017).