Chain-Compatible Differential Pooling
Implementation & benchmark of arXiv:2608.05048 — Differential Homology
Source paper: Differential Homology arXiv:2608.05048 ⓘ · analyzed Aug 31, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Treat learned features on a mesh as differential forms and pool them against oriented chains using wedge or cap products instead of ordinary coordinate averaging. Couple forward and boundary features with the signed chain differential so that pooling commutes with differentiation, preserving local conservation and orientation information.
Formulas
Mathematical statement
The paper constructs two chain maps from \(C_\bullet(X;\mathbb{R})\otimes\Omega_\star(X)\) to currents: \(\gamma_k\otimes\omega_h\mapsto T_{\gamma_k}\wedge\omega_h\) and \(\gamma_k\otimes\omega_h\mapsto T_{\gamma_k\cap\omega_h}\). Their tensor-complex boundary is \(\partial(\gamma_k\otimes\omega_h)=(-1)^h(\partial\gamma_k\otimes\omega_h-\gamma_k\otimes d\omega_h)\), where \(\gamma_k\) is an oriented \(k\)-chain, \(\omega_h\) is a smooth \(h\)-form, \(d\omega_h\) is its exterior derivative, and \(T_{\gamma_k}\) is the current induced by integration over \(\gamma_k\). The chain-map property means that applying the boundary after pooling equals pooling the signed combination of the chain boundary and feature exterior derivative. For a simplicial implementation, represent a 0-form by vertex values, a 1-form by oriented edge values, and a 2-form by oriented face values; the exterior derivative is the coboundary matrix \(D_h\). A practical learned pooling operator is \(P_{k,h}(\gamma,\omega)=\gamma^{\mathsf T}M_{k,h}\omega\), with \(M_{k,h}\) assembled from oriented incidence and local simplex volumes. It should satisfy the discrete commutation residual \(B P_{k,h}\approx(-1)^h[P_{k-1,h}B+P_{k,h+1}D_h]\).
Implementation notes
Insert the module between message-passing layers in a simplicial neural network. The input is a mesh with oriented incidence matrices \(B_k\), dual coboundary matrices \(D_h=B_{h+1}^{\mathsf T}\) under the selected primal/dual convention, simplex features \(\omega_h\), and a chain or attention weight vector \(\gamma_k\). Build sparse local pairing matrices \(M_{k,h}\): for every incident pair of a \(k\)-simplex and an \(h\)-simplex, use the orientation sign multiplied by a normalized overlap or simplex-volume factor. For the MVP, use incidence signs and normalized local volumes rather than exact geometric intersection currents. Compute pool = gamma.T @ M[k,h] @ omega. Also compute chain_pool = (-1)**h * (P[k-1,h](B[k]@gamma, omega) + P[k,h+1](gamma, D[h]@omega)). Add lambda_comm * ||B[out_degree]@pool - chain_pool||^2 to the task loss, or feed this commutator residual into the next-layer MLP as a consistency channel. Keep orientation signs fixed and learn only scalar channel gates initially. The mathematical objects computed exactly are incidence signs, boundary matrices, coboundary matrices, and the signed Leibniz rule; volume normalizations and \(\lambda_{\mathrm{comm}}\) are empirical. Run a cheap ablation on a synthetic triangulated-surface classification or flow-prediction task, comparing mean pooling, graph attention, and differential pooling at equal FLOPs. Evaluate accuracy under random face flips, remeshing, and orientation-preserving coordinate transforms, together with the commutation residual. The expected signal is improved robustness to triangulation and orientation changes and faster convergence when labels depend on circulation, flux, or boundary structure.
Verification
This idea has not been verified yet.
Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.
Artifacts
Artifacts unavailable.