Edgewise Envelopes Between Balanced Forman and Ollivier-Ricci Curvature

Evaluating Ollivier-Ricci (OR) curvature on large-scale graphs is computationally prohibitive due to the necessity of solving an optimal transport problem for every edge. We bypass this computational bottleneck by deriving explicit, two-sided, piecewise-affine transfer moduli between the transport-based OR curvature and the combinatorial Balanced Forman (BF) curvature introduced by Topping et al. By constructing a lazy transport envelope and augmenting the Jost and Liu bound with a cross-edge matching statistic, we establish deterministic bounds for $\mathfrak{c}{OR}(i,j)$ parameterized by 2-hop local graph combinatorics. This formulation reduces the edgewise evaluation complexity from an optimal transport linear program to a worst-case $\mathcal{O}(\max{v \in V} \operatorname{deg}(v)^{1.5})$ time, entirely eliminating the reliance on global solvers. We validate these bounds via distributional analyses on canonical random graphs and empirical networks; the derived analytical bands enclose the empirical distributions independent of degree heterogeneity, geometry, or clustering, providing a scalable, computationally efficient framework for statistical network analysis.

💡 Research Summary

The paper tackles a long‑standing computational bottleneck in network geometry: the evaluation of Ollivier‑Ricci (OR) curvature on large graphs. OR curvature is defined via the Wasserstein‑1 distance between the one‑step lazy random‑walk measures at the two endpoints of an edge. Computing this distance requires solving an optimal transport linear program for every edge, which is infeasible for graphs with millions of edges.

To bypass this obstacle, the authors introduce a deterministic “lazy transport envelope” that bounds the OR Wasserstein‑1 cost from above and below using only local combinatorial information. The envelope decomposes a feasible transport plan into three components: (i) zero‑cost matches on the common support (the diagonal), (ii) residual mass that stays at the original vertices (the lazy part), and (iii) unit‑cost flow across cross‑edges that connect a neighbor of one endpoint to a neighbor of the other. By explicitly counting the number of common neighbors (triangles) and the number of cross‑edges that close 4‑cycles, the envelope yields piecewise‑affine upper and lower bounds on the OR cost.

The paper then connects these bounds to the Balanced Forman (BF) curvature, a purely combinatorial edge curvature introduced by Topping et al. BF curvature depends on the degrees of the two endpoints, the triangle count Δ(i,j), and a 4‑cycle coefficient C₄(i,j) that aggregates cross‑edge information. The authors define several comparison parameters: a symmetric degree shift S(i,j)=2deg_i+2deg_j−2, a triangle scaling term T(i,j)=2·deg_max+1/deg_min, a residual curvature factor K(i,j)=1−1/deg_min−1/deg_max, and normalized overlaps Z_min, Z_max. Using these, they write BF curvature as

c_BF(i,j)= S(i,j)+T(i,j)·Δ(i,j)+C₄(i,j)

for edges whose endpoints have degree larger than one.

The central theoretical contribution is a pair of “transfer moduli” φ and ψ (and their inverses) that map a bound on one curvature into a bound on the other. For example, given a lower bound ζ on BF curvature, the lower transfer φ_{BF→OR}(ζ) provides a guaranteed lower bound on OR curvature; similarly ψ_{BF→OR}(ϑ) gives an upper bound. The moduli are piecewise‑affine functions whose breakpoints are determined solely by the local 2‑hop statistics (degrees, Δ, C₄). The authors prove that these moduli are tight: they recover the sharp Jost‑Liu lower bound for OR and improve upon naive degree‑only estimates.

From an algorithmic standpoint, computing the required local statistics for every edge can be done in worst‑case O(max deg^{1.5}) time per edge, because triangle counting and cross‑edge enumeration involve intersecting neighbor lists. Consequently, the entire edgewise OR curvature approximation runs in O(|V|·max deg^{1.5}) time, eliminating the need for global linear‑program solvers. This represents a dramatic speed‑up compared with the naïve O(|E|·LP) approach.

The empirical section validates the theory on a broad suite of synthetic and real‑world networks. Synthetic models include Erdős‑Rényi, Watts‑Strogatz, Barabási‑Albert, random geometric graphs, d‑regular graphs, stochastic block models, and hyperbolic random graphs. Real networks span the Zachary Karate Club, Jazz collaboration network, US power grid, yeast transcription network, and the arXiv HEP‑Ph citation graph. For each dataset the authors compute the exact OR curvature (using an LP solver) and the BF curvature, then plot the derived envelope bands. The bands consistently enclose the empirical OR distribution regardless of degree heterogeneity, clustering coefficient, or underlying geometry. Notably, in highly clustered small‑world graphs the upper envelope rises sharply, reflecting the strong contribution of triangles, while in tree‑like or regular graphs the lower envelope dominates, mirroring the scarcity of overlapping neighborhoods.

Beyond validation, the authors discuss practical implications. The transfer moduli enable rapid OR‑based community detection, curvature‑guided graph neural network features, and curvature‑driven diffusion or robustness analyses without incurring the cost of solving optimal transport problems. Moreover, the deterministic bounds provide a principled way to quantify uncertainty in curvature estimates, which can be useful for statistical hypothesis testing on network geometry.

In summary, the paper makes four major contributions: (1) it derives explicit, two‑sided, piecewise‑affine transfer functions linking OR and BF curvatures; (2) it constructs a lazy transport envelope that yields deterministic edgewise bounds using only 2‑hop combinatorial data; (3) it shows that these bounds can be computed in O(max deg^{1.5}) time per edge, making curvature evaluation scalable to massive graphs; and (4) it empirically demonstrates that the bounds are tight across a wide variety of synthetic and real networks. This work bridges the gap between transport‑based and combinatorial curvature notions, providing a computationally efficient toolkit for curvature‑driven network analysis.