acceptodds
Under review as a conference paper at ICLR 2027

Path Laplacian Encodings for Directed Graphs

Abstract

Directed graphs naturally model many real-world systems in which interactions are asymmetric, such as citation networks, web graphs, and information-flow networks. However, graph learning methods commonly rely on message passing on symmetrised graph representations or positional encodings that only partially exploit edge directionality. We introduce PathLapPE, a novel spectral positional encoding (PE) derived from the path Laplacian on directed graphs. PathLapPE provides both node- and edge-level features that encode directional higher-order structure and can be incorporated into standard graph learning architectures. Empirical results on node- and graph-level benchmark tasks show that PathLapPE yields consistent improvements across several architectures, especially when combined with direction-aware message passing. Compared with Magnetic Laplacian positional encodings, a widely studied spectral positional encoding for directed graphs, PathLapPE does not require additional fine-tuning of directionality hyperparameters while offering competitive runtime and performance.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Positional encodings, graph learning, directed graphs, path Laplacian

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_Un9rOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes PathLapPE, a positional encoding for directed graphs derived from the path Laplacian of the directed path complex. It provides real-valued node- and edge-level features and, unlike magnetic Laplacian encodings, needs no tuning of the charge parameter q. The authors also introduce EdgeSPE, an edge-level extension of SPE that makes the edge encodings sign- and basis-invariant. Experiments on synthetic, node-level and graph-level benchmarks show PathLapPE obtains competitive performance against the other directed (and undirected) graph positional encoding methods.

Critical strengths

The paper is well written, and the related work section gives a helpful overview for readers less familiar with the topic. The motivation for building positional encodings from the path Laplacian of the directed path complex, rather than the magnetic Laplacian, is convincing, since tuning the charge parameter q for each dataset can be costly. The experiments cover a range of settings and show how the method behaves across different benchmarks.

Critical weaknesses

However, the node classification evaluation is limited, as it relies on three small WebKB graphs and Roman-Empire. The paper would benefit from additional directed benchmarks commonly used in the directed-GNN literature. For example, Telegram, used in MagNet (Zhang et al., 2021) and more recently in Fiorini et al. (2026), is a directed influence network on which direction-aware models clearly outperform undirected ones. This makes it a natural test of whether PathLapPE captures useful directional information. Moreover, although Appendix H analyses computational cost, all graphs in the paper are small compared with real-world scenarios, the largest having about 22k nodes. Evaluating on large directed benchmarks such as arxiv-year or snap-patents, as used in Dir-GNN (Rossi et al., 2024), would show how the method compares with the alternative methods on very large graphs, in both accuracy and computational cost.

I also find the motivation for EdgeSPE underdeveloped. The paper notes that SPE cannot be applied directly to edge-level encodings, but I don't find it too fitting in the general narrative. Empirically, based on the experiments, its benefit is not very clear. I believe the paper should frame EdgeSPE more clearly.

References:

Fiorini, S., Aktas, H. E., Duta, I., Morerio, P., Del Bue, A., Lio, P., & Coniglio, S. (2026). Sheaves Reloaded: A Direction Awakening. International Conference on Learning Representations (ICLR).

Rossi, E., Charpentier, B., Di Giovanni, F., Frasca, F., Gunnemann, S., & Bronstein, M. M. (2024). Edge directionality improves learning on heterophilic graphs. Proceedings of the Second Learning on Graphs Conference, PMLR 231, 25:1-25:27.

Zhang, X., He, Y., Brugnone, N., Perlmutter, M., & Hirn, M. (2021). MagNet: A neural network for directed graphs. Advances in Neural Information Processing Systems, 34, 27003-27015.

Questions for authors
  1. Rationale for EdgeSPE. What is the motivation for the specific design of EdgeSPE (diagonal, successor and predecessor terms)? From the experiments, it seems to help mainly on the graph-level tasks (OCB, HLS) and much less on the node-level ones. How does EdgeSPE fit the paper's narrative?

  2. Choice of node classification datasets. Why were only the WebKB datasets and Roman-Empire used for node classification? Did the authors try other directed benchmarks such as Telegram (Zhang et al., 2021), where direction-aware models clearly outperform undirected ones?

  3. Cost on larger graphs. Since L1 grows with the number of edges, could the authors report preprocessing time and memory on a larger directed graph, such as arxiv-year (Rossi et al., 2024)? If the method is not feasible at that scale, this would limit its practical use.

9 comments
AuthorsOpenReview ↗

Rebuttal overview and changes to PDF

We thank reviewer Un9r for their time and valuable feedback. We have aimed to answer their questions and address their concerns, and of course would be happy to address any follow-ups. In particular, we motivate here the introduction of EdgeSPE and provide further details on its construction and invariance properties. We also uploaded a revised PDF, in which new and updated sections are colored blue. We provide several additional baselines in the revised PDF, also marked in blue:

  • Multi-q-MagLapPE with #q = 3, 5, 10 has been added to WebKB (Table 1).
  • Multi-q-MagLapPE with #q = 5, 10 has been added to Roman-Empire (Tables 2 & 6).
  • DirRWPE (a directed random-walk-based PE, based on Huang et al. (2025)) has been added to WebKB (Table 1) and OCB/HLS (Table 3). DirRWPE raises OOM errors on large graphs and therefore is not evaluated on Roman-Empire or arxiv-year.
  • We have also rerun all PathLapPE results using the same hardware & software stack (package versions etc.) for consistency. These results are also updated across our tables, but any differences are negligible and our conclusions remain the same.
AuthorsOpenReview ↗

Q1 & W3 (Part 1): Motivation for EdgeSPE

We agree with the reviewer that the discussion on EdgeSPE should be expanded. We would therefore like to clarify how it fits within our framework, motivate our choice of aggregation, and provide a brief explanation of its invariance properties.

Motivation

Our PE is spectral in nature, meaning that it is basis-dependent and can be prone to instability: small perturbations of the Laplacian may lead to significant changes in its eigenvectors. This motivates processing the PE to obtain stable descriptors that are invariant under changes of basis. That is precisely the purpose of SPE (Huang et al., 2024), which is also essential for the success of Multi-q-MagLapPE on graph-level benchmarks (Huang et al, 2025). In our case, we have to process spectral information from two Laplacians, L0L_0 and L1L_1. The node-level encodings from L0L_0 can be processed by the standard (Node)SPE of Huang et al. (2024). The edge-level encodings from L1L_1 are not covered by it and require an analogous treatment. EdgeSPE provides this treatment by closely following the approach of Huang et al. (2024). Within our framework, it is therefore the edge-level counterpart of SPE.

Benefits on graph-level tasks

Indeed, we expect the gains from EdgeSPE to appear mainly on graph-level benchmarks, as the reviewer observes. On node-level benchmarks, the train/validation/test splits are defined over node subsets of a single graph, and the encodings are computed once and cached. Thus the choice of signs and bases is arbitrary but fixed throughout training, so the model never sees the same structure encoded in two different bases, effectively circumventing sign and basis ambiguity. In contrast, on graph-level benchmarks, eigenvectors need to be computed independently for each graph. Structurally similar edges in different graphs can therefore receive encodings that differ by an arbitrary sign flip or basis. Without EdgeSPE, the model must learn approximate robustness to this variation from data, which effectively injects noise into training. EdgeSPE removes this variation by construction.

This matches our empirical observations. On the OCB & HLS regression tasks, PathLapPE with EdgeSPE consistently outperforms its counterpart without EdgeSPE (Table 3 of the revised paper). The effect is even larger on FlowGraphs, where the task depends exclusively on the directional information carried by the PEs:

PathLapPEFG2FG3FG6
(+) EdgeSPE97.56 $\pm$ 0.4774.25 $\pm$ 2.4738.75 $\pm$ 2.54
(—) EdgeSPE83.88 $\pm$ 2.9249.17 $\pm$ 9.3522.63 $\pm$ 5.47
AuthorsOpenReview ↗

Q1 & W3 (Part 2): Motivation for EdgeSPE

Construction and motivation for the aggregation

Following Huang et al. (2024), we construct sign- and basis-invariant spectral kernels from the eigenvectors and eigenvalues of L1. Let $U\in\mathbb{R}^{|E|\times k}$ contain the eigenvectors of L1 associated with the k smallest eigenvalues $\lambda\in\mathbb{R}^k$. We define the spectral kernels

$$
\quad K_\ell=U\operatorname{diag}(\phi_\ell(\lambda))U^\top,
\qquad \ell=1,\ldots,m,
$$

where the functions $\phi_\ell$ are permutation-equivariant, as in Huang et al. (2024). Since L1 acts on directed edges, the resulting kernels $K_\ell\in\mathbb{R}^{|E|\times|E|}$ are naturally indexed by pairs of directed edges. To obtain an encoding supported on each individual edge, we introduce an aggregation step.

Our choice of aggregation is motivated by the directed path structure underlying PathLapPE. We consider a simple and interpretable construction that combines the diagonal entries of the spectral kernels with their values on predecessor and successor edges. More precisely, for a directed edge e, we consider its sets of successors and predecessors,

$$
\quad S(e)= \{ f:\operatorname{tail}(f)=\operatorname{head}(e) \},
\qquad
\quad P(e)=\{f:\operatorname{head}(f)=\operatorname{tail}(e) \}.
$$

For each kernel $K_\ell$, we compute

$$
\quad h_\ell^{\mathrm{diag}}(e)=K_\ell(e,e), \ \qquad h_\ell^{\mathrm{succ}}(e)=\sum_{f\in S(e)}K_\ell(e,f),\
\qquad h_\ell^{\mathrm{pred}}(e)=\sum_{f\in P(e)}K_\ell(f,e).
$$

The diagonal term captures spectral information associated with the edge itself, while the predecessor and successor terms capture its relationships with edges that can respectively precede or follow it in a directed 2-step walk. By treating these two relationships separately, the aggregation preserves the distinction between the two possible positions of an edge in such a walk. This provides a natural way to obtain edge-level representations from the spectral kernels while retaining information about the directed structure of the graph. Our implementation also allows users to select which of these aggregation terms to retain, providing flexibility in the construction of EdgeSPE.

The resulting features are concatenated and processed by an MLP to obtain the final edge-level encodings.

Invariance properties

The sign- and basis-invariance of EdgeSPE follow directly from the construction of the spectral kernels, using the same arguments as in Huang et al. (2024):

  • Sign invariance: Changing the sign of an eigenvector leaves its contribution to the kernels unchanged, since these depend on the outer products uiui⊤u_i u_{i}^{\top}.
  • Basis invariance: For repeated eigenvalues, the permutation-equivariance of $\phi_\ell$ ensures that the corresponding eigenvectors receive equal weights. The kernels are therefore invariant under orthogonal changes of basis within the associated eigenspaces.
  • Preservation under aggregation: Since the kernels are sign- and basis-invariant, taking their diagonal entries, aggregating over predecessor and successor edges, and applying an MLP preserve these properties.

Thus, EdgeSPE provides sign- and basis-invariant edge-level representations while retaining the directed structural relationships described above.

[Huang et al., 2024] On the Stability of Expressive Positional Encodings for Graphs. ICLR 2024.
[Huang et al., 2025] What are Good Positional Encodings for Directed Graphs? ICLR 2025.

Reviewer_uXwPOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PathLapPE, a positional encoding method based on the path Laplacian of directed graphs. The method uses the spectral information of the zero-order and first-order path Laplacians to construct node features and edge features respectively, in order to describe the structural information in directed graphs, especially the higher-order relations which involve directed 2-paths. The authors give a method to construct an orthonormal basis of the related path space, and they also propose EdgeSPE, which extends the idea of stable spectral encoding to the edge-level representation. The authors evaluate the method on a synthetic directionality task, the WebKB datasets, Roman-Empire, and regression tasks on circuit graphs and program graphs, and compare it with other methods such as the standard Laplacian encoding and the magnetic Laplacian encoding. The experimental results show that PathLapPE is competitive in many settings, and it is especially effective when it is combined with EdgeSPE and direction-aware message passing architectures

Critical strengths
  1. The paper proposes a new real-valued spectral positional encoding for directed graphs based on the path Laplacian. It can provide both node features and edge features, and it does not need to tune a phase parameter like the magnetic Laplacian encoding.
  2. The authors give an explicit construction of the orthonormal basis of the ∂\partial-invariant 2-path space (Theorem 1), which makes the computation of the edge Laplacian efficient. Besides, the proposed EdgeSPE extends the stable spectral encoding to edge level, and it is shown to be helpful especially in the inductive graph-level tasks.
  3. The mathematical part is clearly written. The worked examples (Examples 1–3 and Figure 2) help the readers who are not familiar with path complexes to understand the construction.
Critical weaknesses
  1. The contribution of the 2-path information is not isolated. The edge Laplacian contains both the down Laplacian B1⊤B1B_1^\top B_1 and the up Laplacian B2B2⊤B_2 B_2^\top, but there is no ablation using only the down Laplacian. So it is unclear whether the gain comes from the higher-order path structure, or just from having edge-level encodings. Since the 2-path structure is a main motivation of the paper, this ablation is important.
  2. The improvement over strong PE baselines is modest in several settings. On Roman-Empire, GAT and BiGAT results are within one standard deviation of other PEs. On OCB/HLS, PathLapPE without EdgeSPE is worse than No PE in many tasks, and No PE is still better on DSP even with EdgeSPE.
  3. The left margin of Figure 2 is too wide, and the text in Figure 4 is too small. Please improve the layout and font size.
Questions for authors
  1. Could the authors add an ablation that uses only the down Laplacian B1⊤B1B_1^\top B_1, and compare it with the full edge Laplacian B1⊤B1+B2B2⊤B_1^\top B_1 + B_2 B_2^\top, while keeping EdgeSPE, the encoding dimension and the training setting the same? This would directly show whether the higher-order path term is the main source of the improvement.

  2. For the Multi-q-MagLapPE baseline, did the authors use the same edge-level feature processing as the original method? The training divergence described in Appendix G.4 is quite severe (e.g., 44.98±39.9444.98 \pm 39.94 on OCB BW with GINE). Could the authors explain the differences in implementation and configuration from the original work, and give some diagnosis of the divergence, to confirm that the comparison is not affected by a fixable reproduction issue?

  3. Are all the choices of kk and qq made only based on the validation set? Please clarify how the "Best" results in the tables are selected, and how hyperparameter selection and result aggregation are done across different data splits and random seeds.

  4. The paper compares PathLapPE with Multi-q-MagLapPE mainly in terms of tuning, computational cost and empirical results, and the discussion of expressive power is mostly qualitative. Huang et al. (2025) show that Multi-q-MagLapPE can represent directed walk profiles. Can PathLapPE represent such information, or are there directed structures that one can distinguish but the other cannot? A theoretical analysis or some controlled examples would make the comparison more convincing

11 comments
AuthorsOpenReview ↗

Rebuttal overview and changes to PDF

We thank reviewer uXwP for their time and valuable feedback. We have aimed to answer their questions and address their concerns, and of course would be happy to address any follow-ups. We have uploaded a revised PDF, in which new and updated sections are colored blue. We also provide several additional baselines in the revised PDF, also marked in blue:

  • Multi-q-MagLapPE with #q=3,5,10 have been added to WebKB (Table 1)
  • Multi-q-MagLapPE with #q=5,10 have been added to Roman-Empire (Tables 2 & 6)
  • DirRWPE (a directed random-walk based PE, based on Huang et al., (2025)) has been added to WebKB (Table 1) and OCB/HLS (Table 3). DirRWPE raises OOM on large graphs, and therefore is not evaluated on Roman-Empire.
  • We have also rerun all PathLapPE results using the same hardware & software stack (package versions etc.) for consistency. These results are also updated across our tables, but any differences are negligible and our conclusions remain the same.
AuthorsOpenReview ↗

Q1 & W1 (Part 1): Higher order directional information in

Summary

We thank the reviewer for raising this concern. We believe that two related but distinct questions should be considered: (1) the contribution of node-level versus edge-level spectral information, and (2) the contribution of the up component of the 1-path Laplacian, which explicitly incorporates admissible 2-paths. For (1), we conducted an ablation on Roman-Empire showing that all the bidirectional models benefit from the additional edge encodings. For (2), we conducted another ablation on the same dataset, which did not establish a clear advantage for the full Laplacian over its down component. Additionally, we provide an example illustrating how the full edge Laplacian can distinguish two orientations of the same undirected four-cycle through their spectra, whereas their down Laplacians are isospectral. We also discuss the differences between the directional information captured by the up and down components of L1L_1. We remark that both the up and down Laplacians encode higher-order directional information through different approaches.

AuthorsOpenReview ↗

Q1 & W1 (Part 2): (a) Contribution of node-level versus edge-level encodings.

We first clarify that our framework constructs spectral encodings from two different operators, L0L_0 and L1L_1, supported respectively on nodes and directed edges. The node-level and edge-level operators capture different aspects of the graph structure. In particular, while L0L_0 captures the connectivity structure of the underlying graph, L1L_1 additionally captures relationships between directed edges.

We would also like to emphasize that, in our experimental pipeline, both node-level and edge-level positional encodings are ultimately incorporated into the GNNs as edge features. Node-level spectral information is processed by NodeSPE, while edge-level information is processed by the corresponding edge encoder or EdgeSPE. Consequently, the comparison does not simply contrast models receiving node features with models receiving edge features, but directly compares the information content of the respective PEs.

To isolate the contribution of the node-level L0L_0-Laplacian encodings from the full PathLapPE (L0+L1L_0 +L_1) encodings, we conducted the following ablation on Roman-Empire.

Ablation 1: Node vs. Edge Path Laplacian Encodings

Ablation of the node- and edge-level components of PathLapPE on Roman-Empire. We include LaplacianPE as a node-level reference. The final column reports the best result across the considered values of k; 🥇, 🥈, and 🥉 denote the three best variants for each encoder.

BiGINE-shared

PE variantk=8k=16k=32k=64Best
LaplacianPE87.68 ± 0.4287.44 ± 0.2387.60 ± 0.3187.53 ± 0.4287.68 ± 0.42
PathLapPE (Node only)87.43 ± 0.3187.67 ± 0.4687.52 ± 0.3787.80 ± 0.4187.80 ± 0.41
PathLapPE (256×2)89.34 ± 0.5689.35 ± 0.5289.27 ± 0.3389.36 ± 0.25🥉 89.36 ± 0.25
PathLapPE89.38 ± 0.4089.07 ± 0.3488.97 ± 0.5689.26 ± 0.39🥈 89.38 ± 0.40
PathLapPE (EdgeSPE, 256×2)89.38 ± 0.4989.45 ± 0.3989.34 ± 0.4289.30 ± 0.53🥇 89.45 ± 0.39
PathLapPE (EdgeSPE)89.32 ± 0.4789.15 ± 0.5389.21 ± 0.5389.19 ± 0.4089.32 ± 0.47

BiGINE

PE variantk=8k=16k=32k=64Best
LaplacianPE90.93 ± 0.4490.97 ± 0.4290.84 ± 0.3890.89 ± 0.3190.97 ± 0.42
PathLapPE (Node only)90.98 ± 0.2690.94 ± 0.2090.80 ± 0.4090.85 ± 0.3990.98 ± 0.26
PathLapPE (256×2)91.22 ± 0.3691.42 ± 0.3691.32 ± 0.2491.22 ± 0.28🥈 91.42 ± 0.36
PathLapPE91.32 ± 0.3591.00 ± 0.3591.15 ± 0.3691.36 ± 0.25🥉 91.36 ± 0.25
PathLapPE (EdgeSPE, 256×2)91.11 ± 0.2891.09 ± 0.2891.14 ± 0.2191.01 ± 0.3691.14 ± 0.21
PathLapPE (EdgeSPE)91.35 ± 0.3191.43 ± 0.2391.32 ± 0.2791.36 ± 0.27🥇 91.43 ± 0.23

BiGAT

PE variantk=8k=16k=32k=64Best
LaplacianPE91.15 ± 0.4091.03 ± 0.3990.97 ± 0.5991.06 ± 0.4491.15 ± 0.40
PathLapPE (Node only)91.19 ± 0.4591.06 ± 0.3691.23 ± 0.3691.03 ± 0.39🥈 91.23 ± 0.36
PathLapPE (256×2)91.10 ± 0.3791.01 ± 0.4191.22 ± 0.4391.11 ± 0.50🥉 91.22 ± 0.43
PathLapPE90.88 ± 0.4390.95 ± 0.4191.17 ± 0.5091.02 ± 0.3191.17 ± 0.50
PathLapPE (EdgeSPE, 256×2)91.13 ± 0.4891.11 ± 0.3491.06 ± 0.5591.25 ± 0.64🥇 91.25 ± 0.64
PathLapPE (EdgeSPE)90.92 ± 0.4290.87 ± 0.3291.11 ± 0.4691.03 ± 0.3591.11 ± 0.46

BiGCNE

PE variantk=8k=16k=32k=64Best
LaplacianPE89.31 ± 0.5189.58 ± 0.5789.40 ± 0.4189.46 ± 0.5389.58 ± 0.57
PathLapPE (Node only)89.40 ± 0.4089.67 ± 0.3789.49 ± 0.4489.51 ± 0.4189.67 ± 0.37
PathLapPE (256×2)90.28 ± 0.3090.25 ± 0.4290.14 ± 0.5490.27 ± 0.35🥈 90.28 ± 0.30
PathLapPE90.19 ± 0.3090.20 ± 0.4190.28 ± 0.4390.20 ± 0.35🥉 90.28 ± 0.43
PathLapPE (EdgeSPE, 256×2)90.23 ± 0.3690.23 ± 0.4090.22 ± 0.4190.19 ± 0.4590.23 ± 0.36
PathLapPE (EdgeSPE)90.43 ± 0.4490.28 ± 0.3890.32 ± 0.4190.43 ± 0.33🥇 90.43 ± 0.44

The table shows that PathLapPE with the spectral encodings obtained from L1L_1 performs consistently better than PathLapPE (Node only), where we consider only the node encodings obtained from L0L_0, on all bidirectional GNN architectures. We can also remark that the L0L_0 Laplacian encodings perform similarly to the classical Laplacian encodings. This is not surprising given the similarity of the two operators, as mentioned in Section 3.2 of our paper.

Reviewer_V1rVOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces PathLaplacianPE which are spectral encodings of edges on directed graphs. The theory of path laplacian provides a theoretically-grounded way of encoding information regarding directional edges of a graph, and the authors use it as an encoding of that edge which is then passed as edge features to neural networks. The method involves computing the "1-path Laplacian" for which this paper also presents an efficient algorithm. The eigenvectors and eigenvalues of this E x E matrix are then used to form the edge features that will be incorporated appropriately into GNNs. The experiments are done on a variety of datasets (synthetic FlowGraph, WebKB, Roman-Empire, OCB, HLS), which increases confidence that the method generally works.

Critical strengths
  • The paper nicely connects GNNs to the theory of path Laplacians.

  • The paper substantiates its claims with a broad range of experimental evaluations.

  • The method performs similar or better to Multi-q-LapPE while having fewer hyperparameters.

Critical weaknesses

No weaknesses that are critical to suggest rejection at this point.

Questions for authors
  • There is no code provided in the supplementary material of this submission. Can the authors confirm that code will be released upon publication of the paper.I would still appreciate the authors uploading code in this submission (if openreview allows) as it helps clarify certain implementation details.

  • Near line 241, the authors state that Theorem 1 assumes the graph does not have any bi-directional edges. But in real world graphs (even the ones considered in this paper), the bidirectional edge motif is present. How does the path laplacian computation algorithm side-step this restriction?

2 comments

We thank reviewer V1rV for their positive and thoughtful feedback. We have aimed to answer their questions below, and of course would be happy to address any follow-ups. We have uploaded a revised PDF, in which new and updated sections are colored blue. We also provide several additional baselines in the revised PDF, also marked in blue:

  • Multi-q-MagLapPE with #q=3,5,10 have been added to WebKB (Table 1)
  • Multi-q-MagLapPE with #q=5,10 have been added to Roman-Empire (Tables 2 & 6)
  • DirRWPE (a directed random-walk based PE, based on Huang et al., (2025)) has been added to WebKB (Table 1) and OCB/HLS (Table 3). DirRWPE raises OOM on large graphs, and therefore is not evaluated on Roman-Empire.
  • We have also rerun all PathLapPE results using the same hardware & software stack (package versions etc.) for consistency. These results are also updated across our tables, but any differences are negligible and our conclusions remain the same.

Q1

We thank reviewer V1rV for their comment, and indeed confirm that our code will be made available alongside the camera-ready version of the paper. If the reviewer has specific concerns or questions regarding the implementation, we would be happy to address them.

Q2

We thank the reviewer for their remark, and apologize for the confusion. Indeed, for real world applications considering double edges is needed, and we did take these into account in experiments, a fact that we should have made clear. Our original proof of Theorem 1 omitted this case, but we have now addressed it properly and added a remark (please see Appendix C). This doesn’t affect the statement of the theorem because the bidirectional edges are ∂\partial-invariant elementary 2-paths which were included as basis vectors.

I thank the reviewers for addressing my questions and updating the proof of Theorem 1 to address bidirectional edges.

Comments on the paper

1 comment

Dear Reviewers,

Thank you for your time and effort in reviewing this paper. The authors have posted responses to your reviews. If you have not yet done so, please read their responses and leave a follow-up comment visible to the authors, indicating which concerns have been addressed and which remain unresolved.

Please also update your review and rating if your assessment has changed. Your feedback will help the authors address any remaining questions during the discussion period.

Thank you!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.