# Divisibility-Weighted Simplicial Message Passing

- ID: 2801
- Canonical URL: https://synthcore.org/idea/2801/divisibility-weighted-simplicial-message-passing
- API JSON: https://synthcore.org/api/idea/2801.json
- API Markdown: https://synthcore.org/api/idea/2801.md
- Verification status: unverified
- Source: [arXiv:2608.29013](https://arxiv.org/abs/2608.29013)
- Category: architecture
- Solves: accuracy, stability, generalization
- ML areas: graph-nn, architecture, regularization
- Math tags: topology, algebra, combinatorics
- Ratings: usefulness 5/10; difficulty 5/10; novelty 6/10

## Idea description

Replace ordinary simplicial incidence matrices in a graph or mesh neural network by integer-ratio weighted incidences derived from a divisibility hierarchy on simplex weights. The resulting up/down message-passing operators preserve exact chain cancellation, so features propagated around a filled simplex cannot create spurious boundary signals. Train the weights either from known metadata or as positive integer powers of a small prime, while retaining an ordinary-incidence baseline for ablation.

## Mathematical statement

A weighted simplicial complex is a pair \((K,w)\), where \(K\) is a simplicial complex and \(w(\sigma)\in\mathbb{Z}_{>0}\) satisfies \(\sigma'\subseteq\sigma\Rightarrow w(\sigma')\mid w(\sigma)\). For an oriented \(n\)-simplex \(\sigma\), the weighted boundary is \(\partial_n^w\sigma=\sum_{i=0}^{n}(-1)^i\frac{w(\sigma)}{w(\partial_i\sigma)}\partial_i\sigma\), where \(\partial_i\sigma\) is its \(i\)-th oriented face; divisibility makes every coefficient an integer. The construction preserves the chain-complex cancellation property \(\partial_{n-1}^w\partial_n^w=0\). The paper also defines a weight-ratio map \(\eta(\theta_0)=\frac{\xi(\Delta^0)}{\mu(\sigma)}\sigma\) in the ascending case and \(\eta(\theta_0)=\frac{\mu(\sigma)}{\xi(\Delta^0)}\sigma\) in the descending case, with W-continuity guaranteeing the relevant divisibility. For a neural layer, let \(B_n^w\) be the matrix representation of \(\partial_n^w\), \(x_n\) features on oriented \(n\)-simplices, and \(\phi\) a pointwise nonlinearity; use \(m_{n-1}=B_n^w x_n\) and \(m_{n+1}=(B_{n+1}^w)^\top x_{n+1}\).

## Key formulas

- $$\sigma'\subseteq\sigma\ \Longrightarrow\ w(\sigma')\mid w(\sigma),\qquad w(\sigma)\in\mathbb{Z}_{>0}.$$
- $$\partial_n^w\sigma=\sum_{i=0}^{n}(-1)^i\frac{w(\sigma)}{w(\partial_i\sigma)}\,\partial_i\sigma,\qquad \partial_{n-1}^w\partial_n^w=0.$$
- $$\eta(\theta_0)=\begin{cases}\displaystyle\frac{\xi(\Delta^0)}{\mu(\sigma)}\sigma,&\text{if ascending type};\\[4pt]\displaystyle\frac{\mu(\sigma)}{\xi(\Delta^0)}\sigma,&\text{if descending type},\end{cases}$$
- $$h_n=\phi\!\left(x_nW_n+\alpha_n(B_n^w)^{\top}x_{n-1}+\beta_nB_{n+1}^wx_{n+1}\right),\qquad \mathcal{L}_{\mathrm{chain}}=\left\|B_{n-1}^wB_n^w\right\|_F^2.$$

## Implementation notes

(1) Integration point: modify the sparse incidence-based message-passing operation in a simplicial GNN or mesh network. Construct oriented vertices, edges, and triangles from every input graph or triangular mesh, and maintain feature tensors \(x_0,x_1,x_2\) for the three simplex dimensions. (2) Pseudocode: assign positive integer weights to all simplices; for every oriented face-coface pair set `B[n][face, coface] = orientation_sign * (w[coface] // w[face])`; compute `down_n = B[n] @ x_n`, `up_n = B[n+1].T @ x[n+1]`, and `x_n = MLP_n(concat(x_n, down_n, up_n))`. Add residual connections and normalize messages by weighted row degree only after constructing the exact integer matrices. (3) Mathematics versus estimation: the ratio \(w(\text{coface})/w(\text{face})\) and the identity \(B_{n-1}^wB_n^w=0\) come directly from the weighted boundary construction. Initially choose weights deterministically as \(w(\sigma)=2^{r(\sigma)}\), with exponents \(r\in\{0,1,2\}\) satisfying \(r(\text{face})\le r(\text{coface})\). If metadata are unavailable, assign exponents from node or element confidence and project them onto this monotonic constraint with a topological pass. Compute `chain_error = norm(B[n-1] @ B[n])`; it must be zero before message normalization. Keep weights fixed in the first experiment so the architectural effect is isolated; only later make exponents learnable using logits followed by integer projection. (4) First experiment: use a 3-layer simplicial GNN on ZINC or a mesh-classification dataset, comparing ordinary incidence matrices with weighted incidence matrices at matched parameter count and FLOPs. Create a synthetic split in which repeated motifs have different simplex multiplicities or confidence levels. Train five seeds with the same optimizer and schedule. Measure validation accuracy, calibration, activation norms, gradient variance, and loss curves. The expected signal is higher accuracy or lower calibration error on heterogeneous-weight examples, with no increase in exploding activations and exactly zero chain error. Also test unweighted random graphs to verify that weighting does not harm the ordinary case.

## Disclaimer

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