# Piecewise-Curvature Message Passing

- ID: 1692
- Canonical URL: https://synthcore.org/idea/1692/piecewise-curvature-message-passing
- API JSON: https://synthcore.org/api/idea/1692.json
- API Markdown: https://synthcore.org/api/idea/1692.md
- Verification status: unverified
- Source: [arXiv:2608.01970](https://arxiv.org/abs/2608.01970)
- Category: geometry
- Solves: accuracy, stability, generalization
- ML areas: graph-nn, embedding, regularization
- Math tags: optimal-transport, metric-geometry, graph-theory, convex-analysis
- Ratings: usefulness 5/10; difficulty 6/10; novelty 6/10

## Idea description

Use local Ollivier–Ricci curvature as a data-dependent controller for the self-loop versus neighbor-mixing coefficient in a graph or hypergraph neural layer. Estimate the idleness-curvature curve from only a few idleness values, then choose a conservative mixing coefficient: highly positively curved edges receive stronger neighbor aggregation, while negatively curved edges retain more self-information to reduce oversmoothing and heterophily damage.

## Mathematical statement

For a locally finite graph G=(V,E), let d_x be the degree of x, let m_x^alpha be the lazy one-step distribution with idleness alpha in [0,1], and let W_1 be the 1-Wasserstein distance under shortest-path distance d. Ollivier curvature on an adjacent pair x~y is kappa_alpha(x,y)=1-W_1(m_x^alpha,m_y^alpha)/d(x,y). The paper's Theorem 1.1 states that alpha maps to kappa_alpha(x,y) is concave and piecewise affine with at most three affine pieces, and is affine on the interval [0,1/(lcm(d_x,d_y)+1)]. For an r-uniform linear hypergraph, the hypergraph walk agrees exactly with the simple random walk on its 2-section graph, whose vertices are V and whose edges connect two vertices occurring in a common hyperedge. We exploit concavity and the three-piece bound to use a small number of curvature evaluations rather than dense idleness search. The neural layer uses an edge-dependent self-loop coefficient alpha_xy; this is distinct from the walk's probability distribution but is chosen from the same geometric signal.

## Key formulas

- $$\kappa_\alpha(x,y)=1-\frac{W_1\!\left(m_x^\alpha,m_y^\alpha\right)}{d(x,y)},\qquad x\sim y,$$
- $$\alpha\longmapsto\kappa_\alpha(x,y)\text{ is concave and piecewise affine on }[0,1]\text{ with at most three affine pieces},$$
- $$\kappa_\alpha(x,y)\text{ is affine on }\left[0,\frac{1}{\operatorname{lcm}(d_x,d_y)+1}\right],$$
- $$h_x^{(\ell+1)}=\sigma\!\left(W_0^{(\ell)}h_x^{(\ell)}+\sum_{y\in N(x)}w_{xy}\,\bigl[(1-\alpha_{xy})h_y^{(\ell)}+\alpha_{xy}h_x^{(\ell)}\bigr]W_1^{(\ell)}\right),\quad \alpha_{xy}=\operatorname{clip}\!\left(\frac12-\beta\,\widehat\kappa_{xy},0,1\right).$$

## Implementation notes

Integrate this into a GCN, GraphSAGE, or hypergraph-to-2-section message-passing layer. For an r-uniform linear hypergraph, first construct the 2-section graph; for a general graph use its ordinary adjacency. At each training epoch or every K epochs, sample a small set of idlenesses A={0,1/2,1}; optionally add the degree-dependent point 1/(lcm(d_x,d_y)+1) for low-degree edges. For each edge (x,y), construct lazy distributions m_x^alpha and m_y^alpha, solve the local optimal-transport problem for W1 using a small ego-graph shortest-path matrix, and compute kappa_alpha=1-W1/d(x,y). Fit a concave piecewise-linear interpolant through these samples, enforcing nonincreasing slopes; use its value at alpha=1/2 as the local curvature estimate kappa_hat_xy. Store this estimate detached from backpropagation. In the layer, compute alpha_xy=clip(1/2-beta*kappa_hat_xy,0,1), where beta is a small positive hyperparameter or a learned nonnegative scalar. Positive curvature increases neighbor mixing, whereas negative curvature increases the self component. Normalize w_xy with standard degree normalization, and compare against the same network with fixed alpha=0, 1/2, or a learned global coefficient. The mathematical guarantee supplies concavity and at most three regimes; the OT values and shortest-path distances are estimated numerically. First test on Cora, Citeseer, and a synthetic heterophilic graph, plus a 3-uniform linear-hypergraph benchmark. Measure validation accuracy, feature oversmoothing via average pairwise cosine similarity, and robustness to depth. Success means improved accuracy or deeper stable training at equal parameter count, with especially visible gains on heterophilic edges and fewer accuracy collapses as layers increase.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.