Unverified 2026

Stable directed-path topology features

Implementation & benchmark of arXiv:2607.26226 — Stability of persistent path homology of path complexes

Usefulness5/10

Difficulty7/10

Novelty6/10

Geometry Accuracy Stability Generalization Graph nn Embedding Attention Regularization Topology Algebraic topology Combinatorics Metric geometry

Source paper: Stability of persistent path homology of path complexes arXiv:2607.26226 ⓘ · analyzed Aug 31, 2026

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 Not peer-reviewed.

Idea description

Construct a filtration from learned directed edge or transition weights, compute persistent path homology, and feed compact persistence features into a graph or sequence neural network. Because the paper proves stability under network-distance perturbations, these features should be less sensitive to small changes in edge scores than raw adjacency statistics, while retaining orientation-sensitive information that ordinary undirected topology loses.

Formulas

$$\partial_n(x_0,\dots,x_n)=\sum_{i=0}^{n}(-1)^i(x_0,\dots,\widehat{x_i},\dots,x_n)$$

$$H_n(P;R)=\ker(\partial_n)/\operatorname{im}(\partial_{n+1})$$

$$H_k^{\mathrm{Path}}(P)=\{H_k^{\mathrm{Path}}(P^\delta)\}_{\delta\geq 0},\qquad d_{\mathrm{path}}(P^q(\overline{G}),P^q(\overline{H}))\leq d_{\mathrm H}(G,H)$$

$$w_{P^q(\overline{G})}(p)=\sup_i\overline{w}_G(b_i),\qquad b_i=(x_i,\dots,x_{i+q-1})$$

Mathematical statement

For a path complex P, let C_n(X;R) be the free R-module generated by regular directed n-paths, obtained from all paths by quotienting irregular paths: C_n(X;R)=K_n(X;R)/D_n(X;R). Its boundary is the alternating deletion operator $\partial_n(x_0,\ldots,x_n)=\sum_{i=0}^n(-1)^i(x_0,\ldots,\widehat{x_i},\ldots,x_n)$, and the path-homology group is $H_n(P;R)=\ker(\partial_n)/\operatorname{im}(\partial_{n+1})$. For a filtration $P^\delta$, where $\delta\geq0$ is a path-weight threshold, the degree-k groups and inclusion-induced maps form a persistence module $H_k^{\mathrm{Path}}(P)=\{H_k^{\mathrm{Path}}(P^\delta)\}_{\delta\geq0}$. An $\varepsilon$-interleaving consists of maps $\varphi_\delta:V_\delta\to W_{\delta+\varepsilon}$ and $\psi_\delta:W_\delta\to V_{\delta+\varepsilon}$ whose two compositions equal the structure map shifted by $2\varepsilon$. The paper proves the weighted-hypergraph comparison $d_{\mathrm{path}}(P^q(\overline G),P^q(\overline H))\leq d_{\mathrm H}(G,H)$, with the induced persistence modules controlled by the corresponding path distortion. Here $R=\mathbb F_2$ is a practical coefficient field, $q$ is the directed sequence-block length, $\overline G$ is the induced weighted path complex, and $d_{\mathrm H}$ is the weighted hypergraph correspondence distance. The useful property is stability of the full persistent descriptor under bounded changes in directed path weights.

Implementation notes

Integrate this at the graph-readout or sequence-embedding stage, using directed edge probabilities or attention transition scores as filtration weights. For each example, construct a sparse directed graph with vertices equal to tokens or nodes and define $w(u,v)=1-a_{uv}$, where $a_{uv}\in[0,1]$ is the normalized learned transition score; low thresholds then add stronger edges first. Choose $q=2$ or $q=3$ and use 16 fixed quantile thresholds. At each threshold $\delta$, retain directed q-paths whose block weight is at most $\delta$, close the collection under endpoint deletion, construct boundary matrices over $\mathbb F_2$, and compute $H_0$ and $H_1$ using sparse Gaussian elimination. Track births and deaths across thresholds through inclusion maps, then convert each interval $[b,d)$ into a fixed persistence-image vector by accumulating a Gaussian kernel centered at $((b+d)/2,d-b)$. Concatenate persistence images with the ordinary pooled embedding and pass them through a small MLP. Pseudocode: A=softmax(scores); E={(u,v):A[u,v]>tau}; for delta in thresholds: P_delta=paths(E,w<=delta); B_n=boundary(P_delta,n); H_n=ker(B_n)/im(B_{n+1}); intervals=match_across_delta(H_n); z=MLP(concat(pool(x),persistence_image(intervals))). The paper supplies the boundary, filtration, and stability construction; estimate the empirical perturbation bound by taking the maximum absolute change in corresponding q-block weights between an original and perturbed graph. Initially detach persistence features from backpropagation to avoid implementing differentiable rank computations. Use a 2-layer directed GAT or GraphSAGE model on Cora, OGB-Arxiv, or a synthetic directed-cycle classification dataset. Compare against the same model with ordinary pooling, degree statistics, and undirected persistent homology. Success means improved test accuracy or robustness to edge deletion and reweighting at equal parameter count, with lower prediction variance under perturbations. Also test the mathematical signal directly: persistence-feature distance should remain bounded by, or at least strongly correlate with, the maximum q-block weight perturbation.

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.