# Volume-Weighted Hodge Convolution

- ID: 1117
- Canonical URL: https://synthcore.org/idea/1117/volume-weighted-hodge-convolution
- API JSON: https://synthcore.org/api/idea/1117.json
- API Markdown: https://synthcore.org/api/idea/1117.md
- Verification status: unverified
- Source: [arXiv:2607.18692](https://arxiv.org/abs/2607.18692)
- Category: architecture
- Solves: accuracy, stability, generalization
- ML areas: graph-nn, geometry, training
- Math tags: topology, linear-algebra, geometry, spectral-theory
- Ratings: usefulness 6/10; difficulty 5/10; novelty 6/10

## Idea description

Replace the ordinary combinatorial Hodge propagation in a simplicial neural network with a geometry-induced weighted Hodge Laplacian built from Euclidean simplex volumes. The operator preserves the harmonic/topological subspace while changing the positive spectrum according to the shape and scale of the simplices, allowing message passing to distinguish geometrically meaningful cells that have identical incidence patterns.

## Mathematical statement

For an oriented simplicial complex K, let B_k be the signed boundary matrix mapping k-simplices to (k-1)-simplices, and let W_k=diag{w_k(σ):σ∈K_k} be a positive diagonal matrix assigning weight w_k(σ)>0 to each k-simplex σ. The paper's weighted Hodge Laplacian on k-cochains is $$L_k^W=W_k^{-1}B_k^T W_{k-1}B_k+B_{k+1}W_{k+1}^{-1}B_{k+1}^T W_k.$$ The first term is the weighted down-Laplacian and the second is the weighted up-Laplacian; omit a term when the relevant boundary matrix does not exist. The weighted inner product is ⟨u,v⟩_{W_k}=u^T W_kv, so L_k^W is self-adjoint in this inner product. For every choice of strictly positive weights, ker(L_k^W) has dimension β_k, the kth Betti number, while the nonzero eigenvalues change with the weights. For the volume star, define w_k(σ)=Vol_k(σ), where Vol_k is the Euclidean k-dimensional simplex volume; in coordinates σ=(x_0,…,x_k), $$Vol_k(σ)=\frac{1}{k!}\sqrt{\det\left((x_i-x_0)^T(x_j-x_0)\right)_{i,j=1}^{k}}.$$ For numerical stability use w_k(σ)=max(Vol_k(σ),ε_w), optionally normalized by the median weight at each dimension. A neural layer can use a scaled polynomial filter of the adapted operator, for example $$H^{(\ell+1)}=\phi\left(\sum_{r=0}^{R}\Theta_r(\widetilde L_k^W)^r H^{(\ell)}\right),\qquad \widetilde L_k^W=\frac{2L_k^W}{\lambda_{\max}}-I,$$ where H^(ℓ) has one row per k-simplex, Θ_r are learnable channel-mixing matrices, φ is an activation, and λ_max is an estimated largest eigenvalue.

## Key formulas

- $$L_k^W=W_k^{-1}B_k^T W_{k-1}B_k+B_{k+1}W_{k+1}^{-1}B_{k+1}^T W_k.$$
- $$\langle u,v\rangle_{W_k}=u^T W_kv,\qquad W_k=\operatorname{diag}\{w_k(\sigma):\sigma\in K_k\},\quad w_k(\sigma)>0.$$
- $$\dim\ker(L_k^W)=\beta_k.$$
- $$Vol_k(\sigma)=\frac{1}{k!}\sqrt{\det\left((x_i-x_0)^T(x_j-x_0)\right)_{i,j=1}^{k}},\qquad H^{(\ell+1)}=\phi\left(\sum_{r=0}^{R}\Theta_r(\widetilde L_k^W)^rH^{(\ell)}\right).$$

## Implementation notes

Integrate this at the simplicial message-passing layer for one chosen cochain degree k, initially k=1 for edge features or k=2 for face features. Construct a Rips complex from point coordinates X using the same radius and maximum simplex dimension as the baseline model. Orient every simplex once and build sparse signed incidence matrices B_k. For each k-simplex, compute its Gram matrix G_ij=(x_i-x_0)^T(x_j-x_0) and set w_k=sqrt(max(det(G),0))/k!; clamp to ε_w, then divide all weights of a fixed dimension by their median. Form sparse W_k and evaluate the two terms in the paper's formula without materializing dense inverses: multiply by reciprocal diagonal vectors W_k^{-1}. Given feature matrix H with one row per k-simplex, compute Y_down=W_k^{-1} B_k^T(W_{k-1}(B_kH)) and Y_up=B_{k+1}(W_{k+1}^{-1}(B_{k+1}^T(W_kH))); set L_H=Y_down+Y_up. Use either H_next=MLP([H,L_H]) or a degree-2 filter H_next=φ(Θ_0H+Θ_1L_H+Θ_2L_H(L_H)). Estimate λ_max by 10-step power iteration on L_k^W and use the rescaled operator only if polynomial filtering is used. The mathematical quantities are the incidence matrices, simplex volumes, weighted Laplacian, and its guaranteed Betti-preserving kernel; empirically estimate λ_max and tune ε_w, filter order, and normalization. First test on a small point-cloud classification benchmark such as ModelNet40 subsampled to 256 points, comparing an unweighted Hodge layer, a graph-only layer, and the volume-weighted layer at equal parameter count and FLOPs. Measure accuracy, calibration, training loss variance, gradient norms, and sensitivity to anisotropic coordinate transformations. A useful success signal is a consistent accuracy gain on geometrically distorted or nonuniformly sampled clouds, with no increase in NaNs or optimization instability; a stronger result is improved accuracy at equal number of simplices and a smaller performance drop when the Rips radius changes.

## Disclaimer

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