Source
Melanie Weber, Emil Saucan and Jürgen Jost (2016), “Characterizing Complex Networks with Forman-Ricci Curvature and Associated Geometric Flows”, arXiv:1607.08654v2, Journal of Complex Networks 5 (2017), LIT-020, read in full in NOTE-026. The Supplemental Material was not read.
What was actually shown
Where it comes from. Forman’s combinatorial Bochner–Weitzenböck formula on a weighted cell complex splits the Laplacian as □_p = B_p + F_p, with B_p non-negative and F_p diagonal; on 1-cells the diagonal entry is the Ricci curvature (Eq. 5). A graph is treated as a 1-dimensional regular CW complex: edges have nodes as children and no faces as parents, which removes half the terms of Forman’s general formula (§2.3, pp. 6–7).
The formula (Eq. 11, checked against the render). For an edge e = (v₁, v₂) with weight ω(e) and node weights ω(v): Ric_F(e) = ω(e) · ( ω(v₁)/ω(e) + ω(v₂)/ω(e) − Σ [ ω(v₁)/√(ω(e)ω(e_v₁)) + ω(v₂)/√(ω(e)ω(e_v₂)) ] ), the sum over edges adjacent to e at v₁ and at v₂. With all weights 1 this is 4 − deg(v₁) − deg(v₂).
What follows. Every term involves only the edge, its endpoints and the edges that share an endpoint with it. No triangle, cycle or longer path enters, because there are no 2-cells; the authors defer triangle and clique terms to “a forthcoming article” (p. 7). Ollivier’s curvature, κ_xy = 1 − W(m_x, m_y)/d(x,y), needs an optimal-transport solve per edge.
What could have come out otherwise. This is a derivation, not an experiment. It could have been wrong in the specialisation; the step is short and standard, though the paper does not write it out and quotes Forman’s Eq. 9 rather than deriving it.
What this does not say
- That Forman and Ollivier curvature agree. The paper says there are “strong theoretical indications” that they are “closely correlated” and defers the evidence to a forthcoming study (C8).
- That the curvature is independent of degree. On a 1-complex it is essentially a degree-and-weight statistic, which the authors concede only in part (p. 8: “Even if curvature cannot be assumed independent of node degrees…”).
- That it applies to arbitrary data weights without error. The formula is derived for standard weights; the paper appeals to Forman’s theorem that admissible weights are dense, and asserts the error is not perceptible (C2).
- That the flows work. The Forman–Ricci flow γ̃ − γ = −Ric_F·γ (Eq. 36) is shown on one Gnutella snapshot pair with no ground truth, and the Laplacian-flow denoising on a schematic karate-club figure. The §4.3 workflow drops the weight factor and so runs the 3-D form the paper had rejected (Eq. 38). The authors say the applications are “only theoretically proposed and illustrated solely on a small example” (§5).
- That real networks look like Barabási–Albert graphs in general. Table 1 shows three real networks closer to one BA sample than to ER or WS samples, and closer to one another in one case (Facebook–BioGrid 0.0086) than to BA.
- That the default weights are consistent. Eq. 21’s signed edge weights contradict Eq. 20’s range and Forman’s positivity requirement.
Connections
- Ollivier–Ricci flow (THEORY-456, from LIT-004). Ni et al. run a flow of the same shape, curvature times current length, with Ollivier curvature, and benchmark it. They report, without data, that Forman curvature did worse for community detection. Neither paper tests a Forman flow on communities.
- Higher-order networks (THEORY-598, from LIT-091). The simplicial complexes and Hodge Laplacians of Battiston et al.’s review are Forman’s full setting; Eq. 11 is its 1-skeleton case, the reading NOTE-057 records. The augmented Forman curvatures with triangle terms, used in later graph-learning work, are not this formula.