Pith. sign in

REVIEW 17 references

VDW-GNNs: Vector diffusion wavelets for geometric graph neural networks

T0 review · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Vector diffusion wavelets extend geometric scattering to vector-valued node signals and provably preserve rotational symmetries, enabling lightweight equivariant GNNs.

desk verdict The vector diffusion wavelet idea is new and the empirical parameter efficiency is real, but Appendix C's sign-flipping claim is false, so Q — and with it Theorems 3.1–3.3 — are not well-defined as written. read the letter →

arxiv 2510.01022 v3 pith:OPSDMS3T submitted 2025-10-01 cs.LG eess.SPstat.ML

David R. Johnson , Alexander Sietsema , Rishabh Anand , Deanna Needell , Smita Krishnaswamy , Michael Perlmutter This is my paper
classification cs.LGeess.SPstat.ML MSC 68T0742C4005C50
keywords vectordiffusionwaveletsgeometricscatteringequivariantgraphneuralnetworksSE(3)-equivariancemapsframeboundspointclouds
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces vector diffusion wavelets (VDWs), a new family of wavelets built from a vector diffusion matrix that acts on vector-valued node features on geometric graphs. The authors prove that these wavelets and their scattering coefficients are equivariant to rotations and translations (SE(3)-equivariance), and that they form a frame with stable inversion bounds. They embed these wavelets into a scattering-based GNN (ESc-GNN) and show on synthetic point-cloud diameter prediction and node-level vector regression that it matches or approaches the accuracy of equivariant message-passing networks like EGNN and TFN while using a fraction—often under 10%—of the parameters. The central claim is that a non-learned, wavelet-based alternative to equivariant message passing can capture global geometric structure and remain rotation-safe.

What carries the argument

The central object is the vector diffusion matrix Q, an nd x nd block matrix whose (i,j) block is P[i,j] O_{i,j}. Here P is the lazy random-walk matrix on the graph, and O_{i,j} = U_i U_j^T is the orthogonal change-of-basis between the d-dimensional local frames U_i, U_j obtained from the signed SVD of each node's relative-distance matrix. Powers Q^m propagate vector features along paths while rotating them to the local frame of the destination node; vector diffusion wavelets are then defined as differences of consecutive powers, mimicking scalar diffusion wavelets and yielding a multiscale, rotation-equivariant filter bank.

What would settle it

Take a 2D two-node graph where node 1's local frame is the identity and node 2's frame is a rotation by 45 degrees: the matrix O = U_1 U_2^T has negative entries, and flipping columns of U_2 cannot make it entrywise nonnegative. Run the paper's sign-flipping algorithm on this graph; if it outputs a Q, then Q differs from the rotation-equivariant parallel transport, and checking whether wavelet outputs under a 90-degree global rotation match the equivariance prediction would reveal the failure. Alternatively, search a random 3D point-cloud k-NN graph for a pair where the sign-alignment rule pro

Watch

Extended reading notes

Core claim

The paper constructs a vector-valued diffusion operator Q on a geometric graph by replacing the scalar random-walk transition probability between neighbors with a product of local coordinate frames: Q[i,j] = P[i,j] O_{i,j}, where O_{i,j} = U_i U_j^T rotates vectors from node j's local SVD basis to node i's. Powers of Q diffuse vector signals while transporting their coordinate frames, and difference operators of these powers (analogous to dyadic diffusion wavelets) yield multiscale, rotation-equivariant filters. The authors prove that these vector diffusion wavelets satisfy frame bounds (Theorem 3.1) and that both the wavelets and their iterated scattering coefficients are SO(d)-equivariant

Load-bearing premise

The construction assumes that the sign-flipping procedure can always align every pair of local SVD frames so that the overlap matrix O_{ij} has only nonnegative entries, but in general no such alignment exists, so the matrix Q on which all theorems and the implementation rely may be ambiguous.

Editorial extensions

If this is right

  • If the equivariance theorems hold, vector-valued signals on geometric graphs can be processed in a rotation-safe way without learned message-passing layers, potentially reducing the parameter complexity of equivariant GNNs by an order of magnitude.
  • The frame bounds imply that vector diffusion wavelet coefficients stably preserve energy, so the transform is invertible and noise-robust in the ℓ2 sense, enabling downstream tasks on the coefficients themselves.
  • Since the wavelets are constructed from graph geometry rather than learned filters, they can be precomputed and cached, making the forward pass of ESc-GNN extremely lightweight compared to standard equivariant models.
  • The generalization to arbitrary dimension d means the method applies to 2D, 3D, and higher-dimensional geometric data, not just molecular or point-cloud settings.
  • The theoretical guarantees extend to the InfoGain-modified wavelet scales used in the experimental architecture, so the practical model inherits the equivariance claims.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's equivariance results rest on the existence of a consistent sign-flipping rule that makes all overlap matrices O_{ij} entrywise nonnegative; a skeptical reader should verify this rule on simple counterexamples—e.g., in 2D, a 45-degree relative rotation between frames—where no column sign flips produce an entrywise nonnegative orthogonal matrix. If the rule fails, the constructed Q is no
  • A direct extension of this work would be to replace the fragile SVD-frame construction with a gauge-independent formulation (e.g., using connection Laplacians or frame bundles), which could preserve equivariance without the sign-alignment assumption and would make the method applicable to graphs with degenerate or repeated singular values.
  • The empirical claims are limited to two synthetic regression tasks; the most informative next test is a real-world molecular property prediction benchmark with diverse edge geometry, where vector features (e.g., velocities, forces) and scarce labels would stress both the equivariance and the parameter efficiency.
  • Since the vector track's invariants (norms and cosine similarities) are rotation-invariant, the architecture implies that all rotation-equivariant information flows through the gated vector combination; a reader might probe whether this bottleneck limits expressivity compared to full tensor-field representations, especially for higher-order angular information.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the vector-wavelet equivariance and frame results are derived from the construction, not fitted or defined in terms of the conclusions; only minor author-overlap citations appear.

full rationale

The claimed derivation chain is not circular. Q is defined by Q[i,j]=P[i,j]O_{ij} with O_{ij}=U_i U_j^T (Section 3), the wavelets are dyadic polynomials of Q (Eq. 3), and Theorems 3.2-3.3 are obtained by proving O_{ij} -> R O_{ij} R^T under a global rotation (Prop. B.3) and the block form Q^m[i,j]=P^m[i,j]O_{ij} (Lemma A.2); these are properties of the construction, not assumptions of the target statements. Theorem 3.1 imports the scalar frame bound (Prop. B.1) from Perlmutter et al. [2023], an author-overlapping citation, but that result is a parameter-free published lemma about scalar diffusion wavelets and does not contain the vector result, so under the review rules it is independent support rather than a circular premise. The InfoGain scale-selection procedure from Johnson et al. [2025b] (Appendix D/E) is an empirical design choice, not a claimed first-principles prediction, and it does not enter the proofs. Non-circular caveats, flagged for correctness rather than circularity: Appendix C's sign-flipping rule is claimed to make every entry of O_{ij} nonnegative and well-defined, but for U_i=I and U_j=R_{π/4} in d=2 no column sign flips make the sign pattern all nonnegative, and flipping U_j per neighbor can break O_{i,k} O_{k,j} = O_{i,j} used in Lemma A.2; if that holds the theorems lack a well-defined operator. Similarly, Appendix D states theoretical guarantees 'may be readily adapted' to the modified wavelet scales without proof. These are gaps, not reductions of outputs to inputs, so the circularity score remains low.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

Central derivation rests on prior scalar frame bounds, SVD local frame construction, distinct singular values, and the (problematic) sign-flip assumption. No new physical entities are introduced.

free parameters (4)
  • Gaussian kernel scale epsilon = squared mean of mean neighbor distances across the final ellipsoid dataset (Appendix E.2.1)
    Used to build B_i = C_i D_i; chosen from the data distribution and affects local frames and Q.
  • InfoGain diffusion scales {t_j} = scalar {0,1,2,4,6,8,16}; vector {0,1,2,4,6,9,16}
    Selected from input features via the InfoGain procedure (Johnson et al. 2025b) on each dataset; controls the wavelet filter bank.
  • Architecture hyperparameters (K, MLP widths, layers, dropout) = K=16 scalar, 32 vector; MLP widths 64/64 and 128/128; readout 128/64/32/16; dropout 0.7 on graph-level task
    Manual architecture choices; the parameter-efficiency claim depends on these counts.
  • Wavelet scale cap t_J and InfoGain quantiles = t_J=16; information cutoff quantiles [0.25, 0.5, 0.75]
    Hand-set hyperparameters that determine the number and placement of wavelets.
assumptions (5)
  • standard math Scalar diffusion wavelets are a nonexpansive frame (Proposition B.1, Perlmutter et al. 2023).
    Imported from prior literature and used in Lemma A.1 to establish vector frame bounds.
  • domain assumption The graph is connected, undirected, weighted, has deg(v_i) >= d, and each B_i has distinct singular values with no multiplicity greater than one.
    Stated in Section 3; needed for the SVD local frame to be unique up to sign flips.
  • domain assumption Rotations preserve the graph edge/weight structure and vector features rotate as R·w; the activation σ commutes with rotations.
    Imposed in Section 3.1 and Theorem 3.3; this is the setting for equivariance but is not automatic for arbitrary geometric graphs.
  • ad hoc to paper Sign-flipping can make every entry of O_ij nonnegative, so O_ij is well-defined.
    Appendix C claims this, but it is false for generic relative rotations; this is the load-bearing gap.
  • ad hoc to paper The InfoGain-modified wavelets inherit the theoretical frame/equivariance guarantees.
    Appendix D asserts this without proof; the experimental network uses InfoGain scales rather than the dyadic scales in the theorems.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VDW-GNNs: Vector diffusion wavelets for geometric graph neural networks." pith.science (2026). https://pith.science/paper/OPSDMS3T

@misc{pith2026251001022,
  author       = {Pith},
  title        = {Pith review of: VDW-GNNs: Vector diffusion wavelets for geometric graph neural networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OPSDMS3T}},
  note         = {Machine review of arXiv:2510.01022}
}
read the original abstract

We introduce vector diffusion wavelets (VDWs), a novel family of wavelets inspired by the vector diffusion maps algorithm that was introduced to analyze data lying in the tangent bundle of a Riemannian manifold. We show that these wavelets may be effectively incorporated into a family of geometric graph neural networks, which we refer to as VDW-GNNs. We demonstrate that such networks are effective on synthetic point cloud data, as well as on real-world data derived from wind field and neural activity measurements. Theoretically, we prove that these new wavelets have desirable frame theoretic properties, similar to traditional diffusion wavelets. Additionally, we prove that these wavelets have useful symmetries with respect to rotations and translations.

Figures

Figures reproduced from arXiv: 2510.01022 by the authors.

Figure 1
Figure 1. Illustration of rotational equivariance of [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Illustration of rotational equivariance of vector diffusion wavelets applied to a vector [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Architecture the equivariant scattering-based GNN. Given a geometric graph, scalar [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 1 canonical work pages

  1. [2]

    Geometric GNN Dojo

    These constructions of P and Q enable highly efficient geometric scattering layers in ESc-GNN, through sparse matrix multiplication of (dense) scalar and vector feature vectors. Still, in terms of complexity of the forward pass of our model, this layer dominates for typical graphs. For each of the scalar and vector tracks, their first-order scattering has...

    2024

  2. [3]

    Within-track mixing layers.Second, for each track, the scattering coefficients are passed through within-track coefficient-mixing layers. For the scalar track, we learn new combinations of the scattering scales using a two-layer MLP, resulting in a new tensor eX′ ∈R n×Fscalar×K defined by eX′[i1, i2,:] =MLP( eX[i1, i2,:]),(14) (where the MLP does not depe...

    2023

  3. [8]

    Manifold filter-combine networks.Sampling Theory, Signal Processing, and Data Analysis, 23(2):17, 2025a

    David R Johnson, Joyce A Chew, Siddharth Viswanath, Edward De Brouwer, Deanna Needell, Smita Krishnaswamy, and Michael Perlmutter. Manifold filter-combine networks.Sampling Theory, Signal Processing, and Data Analysis, 23(2):17, 2025a. David R Johnson, Smita Krishnaswamy, and Michael Perlmutter. Infogain wavelets: Furthering the design of diffusion wavele...

    arXiv

  4. [9]

    Revisiting graph neural networks: All we have is low-pass filters

    Hoang Nt and Takanori Maehara. Revisiting graph neural networks: All we have is low-pass filters. arXiv preprint arXiv:1905.09550,

    arXiv 1905

  5. [10]

    URL https://doi.org/10

    doi: 10.1137/21M1465056. URL https://doi.org/10. 1137/21M1465056. 8 Shaima Qureshi et al. Limits of depth: Over-smoothing and over-squashing in gnns.Big Data Mining and Analytics, 7(1):205–216,

    work page doi:10.1137/21m1465056

  6. [12]

    Convolutional filtering on sampled manifolds

    Zhiyang Wang, Luana Ruiz, and Alejandro Ribeiro. Convolutional filtering on sampled manifolds. InICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE,

    2023

  7. [13]

    Overcoming oversmoothness in graph convolutional networks via hybrid scattering networks.arXiv preprint arXiv:2201.08932,

    Frederik Wenkel, Yimeng Min, Matthew Hirn, Michael Perlmutter, and Guy Wolf. Overcoming oversmoothness in graph convolutional networks via hybrid scattering networks.arXiv preprint arXiv:2201.08932,

    arXiv

  8. [14]

    [2024].) The key to the proof of Proposition B.2 will be to combine this result with the inequality (6), stated below, that relates this weighted norm to the unweightedℓ 2 norm

    and Theorem 1 of Tong et al. [2024].) The key to the proof of Proposition B.2 will be to combine this result with the inequality (6), stated below, that relates this weighted norm to the unweightedℓ 2 norm. Proposition B.1(Proposition 2.2 of Perlmutter et al. [2023]). WJ is a nonexpansive frame, i.e., there exists a universal constant c >0, which in parti...

    2024

Show all 17 references
  1. [16]

    Second, for all models (including ours), we use the sigmoid linear unit (SiLU, or ‘swish’) activation function [Elfwing et al., 2018] in the MLP (except for when another activation function is explicitly prescribed by the baseline model.) Third, when edge weights are used by a...

    2018

  2. [2012]

    Tensor field networks: Rotation- and translation-equivariant neural networks for 3d point clouds

    Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation- and translation-equivariant neural networks for 3d point clouds. arXiv:1802.08219,

    arXiv

  3. [2017]

    Geometric deep learning: Grids, groups, graphs, geodesics, and gauges.arXiv preprint arXiv:2104.13478,

    Michael M Bronstein, Joan Bruna, Taco Cohen, and Petar Veli ˇckovi´c. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges.arXiv preprint arXiv:2104.13478,

    arXiv

  4. [2018]

    Fast graph representation learning with pytorch geometric.arXiv preprint arXiv:1903.02428,

    Matthias Fey and Jan Eric Lenssen. Fast graph representation learning with pytorch geometric.arXiv preprint arXiv:1903.02428,

    1903 arXiv

  5. [2019]

    Geometrically equivariant graph neural networks: A survey.arXiv preprint arXiv:2202.07230,

    Jiaqi Han, Yu Rong, Tingyang Xu, and Wenbing Huang. Geometrically equivariant graph neural networks: A survey.arXiv preprint arXiv:2202.07230,

    arXiv

  6. [2021]

    The manifold scattering transform for high-dimensional point cloud data

    Joyce Chew, Holly Steach, Siddharth Viswanath, Hau-Tieng Wu, Matthew Hirn, Deanna Needell, Matthew D Vesely, Smita Krishnaswamy, and Michael Perlmutter. The manifold scattering transform for high-dimensional point cloud data. InTopological, Algebraic and Geometric Learning Wor...

    2022

  7. [2022]

    Learning graph geometry and topology using dynamical systems based message-passing.arXiv preprint arXiv:2309.09924,

    Dhananjay Bhaskar, Yanlei Zhang, Charles Xu, Xingzhi Sun, Oluwadamilola Fasina, Guy Wolf, Maximilian Nickel, Michael Perlmutter, and Smita Krishnaswamy. Learning graph geometry and topology using dynamical systems based message-passing.arXiv preprint arXiv:2309.09924,

    arXiv

  8. [2023]

    A hitchhiker’s guide to geometric gnns for 3d atomic systems.arXiv preprint arXiv:2312.07511,

    7 Alexandre Duval, Simon V Mathis, Chaitanya K Joshi, Victor Schmidt, Santiago Miret, Fragkiskos D Malliaros, Taco Cohen, Pietro Lio, Yoshua Bengio, and Michael Bronstein. A hitchhiker’s guide to geometric gnns for 3d atomic systems.arXiv preprint arXiv:2312.07511,

    arXiv

  9. [2024]

    doi: https://doi.org/10

    ISSN 1063-5203. doi: https://doi.org/10. 1016/j.acha.2024.101635. URL https://www.sciencedirect.com/science/article/pii/ S1063520324000125. Joyce A Chew, Deanna Needell, and Michael Perlmutter. A convergence rate for manifold neural networks. In2023 International Conference on...

    2024

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.