acceptodds
Under review as a conference paper at ICLR 2027

Circular Topological Transforms: Geometric Chirality with Completeness

14 views

Abstract

Geometric graphs such as point clouds, 3D shapes, and molecular conformations require representations that are both (1) *-complete*, that representations agree if and only if they differ only by translations and rotations, and (2) *chirality-aware*, that distinct geometries and their mirror images remain distinguishable. Existing approaches do not provide both properties at scale. -invariant geometric networks are limited by geometric Weisfeiler–Leman expressivity and capture chirality only locally, while popular topological directional transforms, such as the persistent homology transform (PHT), are injective (can distinguish any shape), but therefore not -complete. In this work, we propose *Circular Topological Transforms* (CIRCUiT), a principled methodology that turns admissible injective directional transforms such as the PHT to be both -complete and chirality-aware on shape classes admitting an equivariant frame. Our construction organizes directional information into ordered circular scans: proper rotations are absorbed by changes of frame, whereas reflections induce a deterministic reversal of the scan. We also show that two naive approaches (group averaging and sorting canonicalization) to symmetrizing directional transforms will fail to detect chirality. Applying our methodology to the PHT, - achieves perfect chirality-sensitive retrieval on multiple molecular benchmarks and captures long-range chirality beyond the interaction radius of the evaluated equivariant deep neural network models.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Directional Transforms, Molecule search, Chirality Detection, Persistent Homology, Topological Data Analysis, Geometric Deep Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_9kW3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Circular Topological Transforms, a way to build geometric representations that are invariant to translations and rotations while still distinguishing a shape from its mirror image. The key idea is to evaluate directional topological descriptors such as the persistent homology transform along ordered circular scans. Experiments on molecular datasets show that the method can distinguish handedness and capture long-range chirality better than several geometric neural-network baselines.

Critical strengths
  1. Clear motivation: the paper explains why standard invariantisation methods can destroy chirality information and formalises this with propositions before introducing the proposed circular transform.

  2. Simple, interpretable construction: the method has an intuitive geometric interpretation. Rotations preserve the ordered circular scan, while reflections reverse it, making the chirality sensitivity easy to understand.

Critical weaknesses
  1. The method depends on the existence and stability of a well-defined equivariant frame. The theoretical guarantees assume an equivariant frame, while the practical construction is equivariant only when both frame-selection stages are defined and return unique directions. An analysis of the robustness of the frame construction to such degeneracies and to small perturbations would strengthen the paper.

  2. Potential computational scalability issue. The construction requires many directional topological computations; for molecular graphs, the stated complexity is roughly O(N^2 log N + DN log N). There is limited empirical evaluation of runtime and scalability relative to modern geometric networks on genuinely large shapes or graphs.

  3. The empirical scope is narrow relative to the general claims: although the method is formulated for general n-dimensional shapes, the experiments are restricted to 3D molecular graphs and are mostly focused on chirality-related tasks. The authors themselves acknowledge that the behaviour of the method on higher-dimensional geometric data remains unexplored.

Questions for authors

I note that the related work does not discuss spectral/topological representations based on the Hodge Laplacian. It would be interesting to compare the proposed method against such approaches [1,2,3], or, if such a comparison is not appropriate, to discuss why. My understanding is that purely combinatorial Hodge-Laplacian signatures would be complementary to the proposed method but would, naively, not distinguish different conformations sharing the same underlying combinatorial structure. If true, this could strengthen the paper’s usefulness.

[1] Graph Classification Gaussian Processes via Spectral Features

[2] Hodgelets: Localized Spectral Representations of Flows on Simplicial Complexes

[3] Graph and Simplicial Complex Prediction Gaussian Process via the Hodgelet Representations

I do not have more questions, except the ones that I raised in the "Critical Weaknesses" section. I am leaning towards rejection at the moment, but addressing the points above could make me raise my score.

5 comments

Thank you for the constructive review and for the pointers to the Hodge-Laplacian literature. We have cited them to better position our work, and we address each point below.

I note that the related work does not discuss spectral/topological representations based on the Hodge Laplacian. [...] purely combinatorial Hodge-Laplacian signatures would be complementary to the proposed method but would, naively, not distinguish different conformations sharing the same underlying combinatorial structure.

Your understanding is right. The Hodge Laplacians of a bond graph depend only on its incidence matrices, so their spectra are the same for a molecule, its rotations, its mirror image and its conformers. We added the references [1, 2, 3] to the related work (Appendix A) and measured it (Appendix G.6, Table 27): the spectrum of the Hodge 1-Laplacian gives 0.555 / 0.483 / 0.463 / 0.511 mAP on QM9-OR / BZR / COX2 / DHFR, against 1.000 for Circ-PHT.

Your comment made us try something we find more interesting: using spectral features as the directional transform inside the circle method. Along a direction, the height filtration gives a nested family of subgraphs, and any spectral quantity of them is a directional transform. We tried three, with the PCA frame and three directions (Tables 28-29):

Directional transformQM9-ORBZRCOX2DHFR
Hodge-1 spectrum, whole graph0.5550.4830.4630.511
Circ-Hodge0.9980.9690.9810.976
Circ-SpectRe1.0001.0001.0001.000
Circ-Hodgelet1.0001.0001.0001.000

Circ-Hodge reads the spectra of L0 and L1 of the sublevel subgraphs. Circ-SpectRe follows SpectRe [4]: each persistence pair carries the Laplacian spectrum of its component. Circ-Hodgelet follows the Hodgelet construction [2, 3] on the sublevel subgraphs, with the vertex heights and edge lengths as signals; since we wanted a training-free descriptor we use fixed wavelet filters where [3] learns them with a Gaussian process. Appendix G.6 gives the exact definitions. So the same spectra that are blind on the whole graph detect chirality once they are organized by the circle method: the method is not tied to persistent homology. Invariance and the reflection law (Theorems 4.4, 4.6) hold for these transforms; completeness needs an injective transform, which spectra are not in general.

[1] Opolka et al., Graph Classification Gaussian Processes via Spectral Features, UAI 2023. [2] Roddenberry et al., Hodgelets: Localized Spectral Representations of Flows on Simplicial Complexes, ICASSP 2022. [3] Alain et al., Graph and Simplicial Complex Prediction Gaussian Process via Hodgelet Representations, NeurIPS 2026. [4] Ji, Souza, Garg, Graph Persistence goes Spectral, NeurIPS 2025.

The method depends on the existence and stability of a well-defined equivariant frame. [...] An analysis of the robustness of the frame construction to such degeneracies and to small perturbations would strengthen the paper.

We have added theoretical and experimental analysis on this.

For Existence: The new Proposition 4.14 proves that an SO(n)-equivariant frame exists generically (proof in Appendix B). In the data, two of the 555 base molecules of the test sets have a proper rotational symmetry (Table 26).

For Stability: The new Theorem 4.15 proves that our transform-guided frame is locally stable: if it is uniquely defined on a geometric graph, it is uniquely defined and varies continuously on all close enough embeddings (Appendix B.5).

The inability to find an equivariant globally stable frame is a well-documented challenge in geometric and equivariant deep learning (see for example [A] & Proposition 3.2 of [B]). In remark B.17, we explain why no equivariant frame can be globally stable. Thus, local stability is the best any researcher could hope for.

We also measured it (Appendix G.1): we add Gaussian noise to the coordinates and compare the frame before and after, for our frame and for a right-handed PCA frame that we added as a second frame function (Appendix E.2.1). At 0.05 Å on BZR the transform-guided frame changes its selected vertices on 71% of the molecules, while 2.6% of the PCA frames rotate by more than 10° (Table 18). Retrieval with three directions at this noise:

DatasetCirc-PHT, transform-guided frameCirc-PHT, PCA frame
BZR0.6620.955
COX20.6440.983
DHFR0.7500.968

On real conformers (3DCS), Circ-PHT with the PCA frame reaches a zero-shot ES-AUC of 0.902 with four directions, against 0.707 in the submission and 0.782 for MolAE (Appendix G.4). The PCA frame has its own degeneracies (repeated eigenvalues, vanishing third moments), where the transform-guided frame is still defined. The circle method takes the frame as an input, so both are covered by the same theorems.

Potential computational scalability issue. [...] There is limited empirical evaluation of runtime and scalability relative to modern geometric networks on genuinely large shapes or graphs.

We measured every stage up to 10,240 vertices on one CPU core (Appendix G.7, Table 30). With the PCA frame the descriptor is linear in the number of vertices: 1.4 ms at 40 vertices, 29 ms at 2,560 and 128 ms with 133 MB of memory at 10,240, with no training and no GPU. The quadratic term belongs to the transform-guided frame (2N directional diagrams), which is why we use the PCA frame for large shapes. On the same graphs one forward pass of EquiformerV2 takes 733 ms and 9.9 GB on a GPU, and it needs training; the spherical model of Paik (2023) needs 492 directions and training and is still not exact (Appendix F.5).

[A] Dym et al. Equivariant frames and the impossibility of continuous canonicalization

[B] Baker et al. An explicit frame construction for normalizing 3D point clouds.

The empirical scope is narrow relative to the general claims: although the method is formulated for general n-dimensional shapes, the experiments are restricted to 3D molecular graphs and are mostly focused on chirality-related tasks.

We added an experiment in R^4 (Appendix G.8). We know of no chirality benchmark in higher dimension, so the data are synthetic, random point sets joined by a sparse graph as in the synthetic experiments of Ruhe et al. [5]. We considered 40 shapes, their mirror images, and 8 random rotations of SO(4) each (640 poses), with the retrieval protocol of Table 1. We use the PCA frame in R^4 and the transform-guided frame extended to R^4, and the scan uses great circles that the reflection reverses. The table is with the PCA frame; with the transform-guided frame Circ-PHT, Circ-ECT and Circ-BCT are also at 1.000 from three directions on.

Directions1234816
Circ-PHT0.5550.5551.0001.0001.0001.000
Circ-ECT0.5550.5551.0001.0001.0001.000
Circ-SpectRe0.5550.5551.0001.0001.0001.000
Circ-Hodgelet0.5550.5551.0001.0001.0001.000
order pooled (blind)0.5550.5550.5550.5550.5550.555
lab frame0.070.080.080.090.090.09

The behavior is the same as in R^3: rotations leave the descriptor unchanged, the mirror image gives the reversed scan, one or two directions are reflection-blind, and three directions separate the two hands (Table 31). The experiments are also broader within molecules now: coordinate noise, conformers, learned and training-free read-outs, spectral transforms, and cost up to 10,240 vertices (Appendix G).

Thank you again for the feedback. The Hodge-Laplacian suggestion in particular led to a new family of circular transforms, and we have cited the suggested works in the revised paper.

[5] Ruhe, Brandstetter, Forré, Clifford Group Equivariant Neural Networks, NeurIPS 2023.

Reviewer_K8L6OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes circular topological transforms for rotation-invariant, chirality-sensitive representations of geometric shapes. It constructs a shape-dependent frame, completes it to a proper rotation, and evaluates PHT, ECT, or BCT along ordered circular scans. Under explicit assumptions, the authors prove rotation invariance, characterize the reflection action, and establish eventual separation of each fixed pair of distinct rotational orbits as sampling is refined. Experiments cover molecular retrieval, long-range chirality, and stereoisomer separation.

Critical strengths

Nice work. The paper is well-written, easy to follow. The insights and utility are clear. The experiments, despite not fully matching the theoretical assumptions, are convincing.

  1. The method provides a concrete, training-free way to turn pose-dependent topological transforms into reusable molecular descriptors. The controlled retrieval results are strong: CIRC-PHT reaches perfect mAP with three final scan directions on four datasets.
  2. The main theoretical arguments are sound under their stated assumptions, and gives a clear explanation of the construction and useful guidance to why it may work empirically.
  3. The experiments include useful mechanism controls. Pooling over scan order removes the chirality signal, while increasing the baseline’s interaction radius restores performance on the long-range task. The additional 3DCS experiment goes beyond exact reflected copies and shows that the descriptor can improve the tested molecular representations.
Critical weaknesses
  1. The practical descriptor appears to not inherit the completeness guarantee of Theorem 4.11. The theorem concerns complete directional transforms and progressively refined sampling, while the experiments use fixed scans and finite readouts. A simple counterexample: take a generic shape X and Y=2X (scaled coordinates by a factor of 2), with the same connectivity. They are not SE(3)-equivalent, but RMS normalization makes their inputs and descriptors identical. This does not contradict the theorem, but shows that the implemented pipeline is not as "complete" as what a reader might expect from the theorem.

  2. The robustness of the frame construction is not sufficiently studied. Near a score tie, the hard argmax in Appendix E.1 can select a different axis after a small coordinate perturbation, changing the scan directions. Near-collinear candidates can also make the frame sensitive. Stability of PH at fixed directions does not guarantee stability of this full pipeline. The 3DCS results are useful, but do not directly establish how the method behaves under controlled geometric fluctuations or noise. This seems an important boundary of its practical applicability, because a physical molecule may have multiple conformers, fluctuating bond lengths and angles, and distorted coordinates from measurement or simulation.

Questions for authors
  1. What is the specific benefit of circular placement? The proof of Theorem 4.11 seems to depend on the frame, injectivity, continuity, and sufficiently rich sampling, rather than the circular arrangement. Could the authors compare against ordinary labelled spherical directions with the same frame, transform, direction count, and output dimension? The order-pooling control shows that retaining directional information matters, but does not isolate the advantage of circles.

  2. Could the authors quantify the efficiency advantage over relevant existing methods, e.g. how are the memory, speed, and scaling with molecule size compared to existing approaches? The three final scan directions still require approximately 2N (N is the number of atoms) additional directional computations to construct the frame. This is a clarification question and will not affect my rating negatively.

  3. How robust is the method to bond-length and bond-angle fluctuations, different conformers of the same stereoisomer, and coordinate noise? An idea may be to measure descriptor changes and frame-switching rates as these perturbations increase, and compare them with the separation between stereoisomers. This would help distinguish useful geometric sensitivity from instability introduced by frame selection.

6 comments

Thank you for the positive and careful review. Below we address your concerns and report the new experiments.

The practical descriptor appears to not inherit the completeness guarantee of Theorem 4.11. [...] A simple counterexample: take a generic shape X and Y=2X [...] RMS normalization makes their inputs and descriptors identical.

Thank you for catching this. The rescaling to unit RMS radius is not part of the method: it was used only for pre-processing QM9-OR, so that its heights fall on the fixed grid of the read-out, and BZR, COX2, DHFR and 3DCS are only centered. The sentence in Appendix E that described it for all datasets was wrong and is corrected, with a remark. We have also clarified that the re-scaling was only done in the pre-processing step, and the completeness guarantee of Theorem 4.11 applies the inputs after they have been pre-processed.

Your example is right for QM9-OR as we ran it: with the rescaling, the descriptors of X and 2X are identical, and without it they differ by 57-82% of their norm. We ran the check you suggest: appending the original RMS radius to the descriptor separates X from 2X again, and the retrieval results do not depend on the rescaling. Circ-PHT with three directions, no noise, mAP (the same with either frame):

CoordinatesQM9-ORBZRCOX2DHFR
no rescaling1.0001.0001.0001.000
unit RMS radius1.0001.0001.0001.000
unit RMS radius, radius appended1.0001.0001.0001.000

The QM9-OR poses are stored at unit RMS radius, so for its first row every molecule was multiplied by its own random factor between 2 and 6. The results do not change because all poses of a molecule and of its mirror image share one scale.

The robustness of the frame construction is not sufficiently studied. Near a score tie, the hard argmax in Appendix E.1 can select a different axis after a small coordinate perturbation [...]

Thank you for your suggestion. We have now analyzed the robustness of our frame both theoretically and empirically.

Theory: We have added a new Theorem 4.15 which proves that the transform-guided frame is locally stable (if it is uniquely constructed on a geometric graph, it is uniquely constructible and varies continuously for all close enough embeddings). The frame is not globally stable, since crossing a tie changes the selected axis.

The inability to find an equivariant globally stable frame is a well-documented challenge in geometric and equivariant deep learning (see for example [1] & Proposition 3.2 of [2]). In remark B.17, we explain why no equivariant frame can be globally stable. Thus, local stability is the best any researcher could hope for.

Experiments (Appendix G.1): we perturb the coordinates with Gaussian noise and compare the frame of each molecule before and after. Median rotation between the two frames / fraction of molecules whose frame rotates by more than 10°:

Dataset, σTransform-guided framePCA frame
BZR, 0.01 Å0.57° / 15%0.10° / 0.4%
BZR, 0.05 Å5.89° / 42%0.49° / 2.6%
COX2, 0.05 Å3.60° / 29%0.41° / 1.1%
DHFR, 0.05 Å3.48° / 33%0.45° / 3.1%

In the squared distance between the frame vectors (Table 13), the transform-guided frame on BZR at 0.01 Å has median 1.2e-4 and mean 0.257, against a mean of 0.014 for the PCA frame. The median is small, as the local stability says, and the jumps are the ties you describe: the selected pair of vertices changes on 37% of the BZR molecules at 0.01 Å. Following the suggestion of Reviewer BPn7 we added a right-handed PCA frame as a second frame function (Appendix E.2.1). It moves in proportion to the noise, and it has its own failure, an axis flip when the noise reaches the eigenvalue gaps (Table 18).

[1] Dym et al. Equivariant frames and the impossibility of continuous canonicalization [2] Baker et al. An explicit frame construction for normalizing 3D point clouds.

How robust is the method to bond-length and bond-angle fluctuations, different conformers of the same stereoisomer, and coordinate noise? An idea may be to measure descriptor changes and frame-switching rates as these perturbations increase, and compare them with the separation between stereoisomers.

Thank you for pointing this. We now added the experiments in Appendix G.1, Tables 14-18. Besides the switching rates above, we report the ratio between the change of the descriptor under noise and its distance to the descriptor of the mirror image: on BZR at 0.05 Å it is 0.359 with the transform-guided frame and 0.115 with the PCA frame. Retrieval with three directions at 0.05 Å (noisy query against the clean gallery, mAP):

DatasetTransform-guided framePCA frame
BZR0.6620.955
COX20.6440.983
DHFR0.7500.968

For real conformers we used 3DCS (Appendix G.4). Retrieving the other conformers of the same stereoisomer:

DescriptormAP
Circ-PHT, transform-guided frame, 88 directions0.424
Circ-PHT, PCA frame, 88 directions0.826
Circ-PHT, PCA frame, 4 directions0.796
MolAE0.505
UniMol0.276

On the 3DCS benchmark itself the PCA-frame descriptor reaches a zero-shot ES-AUC of 0.902 with four directions (0.707 in the submission; 0.782 for MolAE).

What is the specific benefit of circular placement? [...] Could the authors compare against ordinary labelled spherical directions with the same frame, transform, direction count, and output dimension?

Our inspiration for the usage of circles came from Figure 5 and the Motivating Example in Section 4. The proof of completeness did not need the structure of circles, but the circles were added as an inductive bias to better detect chirality (as reflections usually breaks the orientation along the circle, see Figure 5 and also Theorem 4.6).

We have now added Proposition 4.13 to justify our intuition. In the proposition, we prove that the ordered circle method is strictly more expressive than the same directional measurements with their order forgotten (so as a set).

In Table 3 before the rebuttal, we also tested Circ-PHT to Circ-PHT with directions pooled. The latter does an averaging operation and forgets about the order and structure of circles, and performs worse than Circ-PHT.

We also empirically validated that preserving circular order does benefit learned read-outs. Using the same transform-guided frame, PHT landscapes, 12 directions, and 128-dimensional output, we compared circular convolution (readout is a circular CNN) with MLPs on labelled directions and on the flattened circular scan (which does not account the circular structure):

Read-outParametersBZRCOX2DHFR
Circular convolution127k0.7770.7510.818
MLP, labelled directions1.05M0.6890.7270.759
MLP, circular scan1.05M0.6620.6710.749

See Table 22 for more details. All three achieve 1.000 on clean retrieval, but the circular convolution is more robust with approximately eight times fewer parameters. The same ordering holds with only 10% of the training data (Table 22). Since the architectures differ, this is evidence for a useful circular inductive bias as opposed to an isolated effect of direction placement.

Could the authors quantify the efficiency advantage over relevant existing methods [...]? The three final scan directions still require approximately 2N additional directional computations to construct the frame.

Yes, the 2N computations belong to the cost of the transform-guided frame, and they are its quadratic stage. They are not needed with the PCA frame, where the whole descriptor is linear in the number of vertices. We measured it on molecule-like graphs up to 10,240 vertices on one CPU core, with every stage counted (building the complex, the frame, the scan and the read-out; Appendix G.7, Table 30). Ours is Circ-PHT with the PCA frame and three directions; EquiformerV2 is one forward pass:

VerticesOurs (CPU)MemoryEquiformerV2 (GPU)Memory
401.4 ms30 MB23 ms73 MB
6408.3 ms31 MB48 ms627 MB
2,56029 ms45 MB172 ms2.4 GB
10,240128 ms133 MB733 ms9.9 GB

Our descriptor needs no training and no GPU. The spherical model of Paik (2023) needs 492 directional Euler curves per shape (0.23 s at 10,240 vertices) plus a trained network, and still does not reach 1.000 on three of the four datasets (Appendix F.5). EquiformerV2 needs training and a GPU, and on the long-range task it needs 8 to 14 times more edges to see chirality beyond three bonds (Section 5.2). The timings are on different hardware, so we report them with the configurations and not as speedups.

Thank you again, these suggestions helped us separate the general construction from the behavior of one particular frame.

Reviewer_BPn7OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper develops topological shape descriptors that are invariant to rotations while remaining sensitive to reflections, allowing them to distinguish a molecule from its mirror image. The descriptors are based on the persistent homology transform (PHT), Euler characteristic transform (ECT), and a Betti curve transform (BCT). For molecular graphs, this yields the training-free descriptors Circ-PHT, Circ-ECT, and Circ-BCT. On four molecular datasets, these descriptors achieve perfect retrieval scores when distinguishing rotated molecules from their mirrored copies.

Critical strengths
  1. the paper is clearly written
  2. the proof is well drafted
  3. Table 4 shows much better scores than the molecular representations
Critical weaknesses
  1. The method is essentially a canonical pose, and the paper does not compare itself with that idea. A canonical pose means rotating every shape to a fixed standard orientation before computing anything. Using a right-handed frame so that rotations are removed but mirror images are kept is a known idea: frame averaging by Puny et al. at ICLR 2022 and canonicalization by Kaba et al. at ICML 2023 both do this. The paper cites them only as methods that fail, and the authors even say canonicalization cannot give chirality-sensitive transforms, although its own method is a canonicalization. I do see value in packaging this for PHT and ECT and in building the frame from the transform's own output. But the paper should say plainly what it adds over "fix a right-handed pose, then apply the plain transform".
  2. The experiments confirm the theorems but do not show what the method adds. In the retrieval test, the queries are exact rotated copies and the hard negatives are exact mirror images. For these inputs, Theorems 4.4 and 4.6 already guarantee a perfect score for any correct implementation, so the test checks the code more than the method. A trained EquiformerV2 also scores 0.998 to 1.000 there. There is no comparison with a simple right-handed PCA alignment plus the same transform, or with the paper's frame plus a simple non-topological read-out. The frame itself is chosen by picking the highest-scoring vertices, so it can jump when the coordinates change a little. The paper never tests small coordinate noise or different conformers of the same molecule, which is exactly the situation in real molecular data. It also does not say which shapes have such a frame at all.
Questions for authors
  1. Would any result in Section 4 fail if the method were replaced by "rotate to the frame, then read the plain transform at any fixed set of directions"? What do the circles and the transform-based frame add?
  2. Is there a class of shapes and a fixed finite scan for which you can prove the descriptor is complete?
  3. How does Table 1 change with small coordinate noise, or with different conformers of one molecule, for Circ-PHT, for a right-handed PCA pose with the same read-out, and for your frame with a simple non-topological read-out? If Circ-PHT holds up at least as well as the PCA pose, I would move to 6.
  4. On 3DCS, how do versions that ignore reflections score, for example a 2-direction scan?
10 comments

Thank you for the thoughtful and constructive review. Below, we provide point-by-point responses to your comments.

The method is essentially a canonical pose, and the paper does not compare itself with that idea. [...] frame averaging by Puny et al. at ICLR 2022 and canonicalization by Kaba et al. at ICML 2023 both do this. [...] the authors even say canonicalization cannot give chirality-sensitive transforms, although its own method is a canonicalization.

Thank you for raising the comparison with canonicalization and frame averaging. Firstly, we would like to state that these references were substantial to our research, and the naive sorting canonicalization method (Section 3.2) was inspired by the optimization-based canonicalization of Kaba et al. (Eq. 4) [1], where we considered selecting a canonical direction by minimizing a directional score. We want to clarify that our group averaging method [A] (Section 3.1) and naive sorting canonicalization [B] (Section 3.2) are not the exact implementations of Puny et al. [2] and Kaba et al. [1]; moreover, the failure of [A, B] is what motivated us to create the circle method. The reviewer correctly pointed out that our method is also a canonicalization method. We have made this clear, made the statement "canonicalization cannot turn ECT [...] into chirality-sensitive transforms" more precise, and renamed the old sorting canonicalization to best-directions canonicalization to improve clarity. We thank the reviewer for raising this concern, since it helped us improve the positioning of our work; we have updated Section 3 to make this clearer.

What the paper adds to "fix a right-handed pose" is what is done after the pose is fixed. The circle method takes a frame function Fr and a discretization d as separate inputs (Definition 4.2). Theorem 4.4 gives rotation invariance for an equivariant frame, Theorem 4.6 shows that on oriented circles a reflection acts as a fixed reversal of the scan, and Theorem 4.11 gives completeness under refinement of the circles. The new Proposition 4.13 shows that the ordered circle method is strictly more expressive than the same directions with their order forgotten. The frame is a choice: in the revision we use both our transform-guided frame (Definition E.2) and the right-handed PCA frame you suggest (Definition E.3), and the theorems hold for either.

A concrete example of why the pose alone is not the whole story is Paik (2023) [3], which is a related work we evaluate (Appendix F.5). Its reflection-sensitive pooling aligns the transform to a learned template over SO(3), so it also fixes a pose, by alignment. With the same number of directions (12), mAP:

MethodQM9-ORBZRCOX2DHFR
Paik, released0.3540.2920.1940.329
Paik, Sec. 3.50.3890.3100.1800.350
Paik, S²-conv0.7550.7130.5400.771
Circ-PHT1.0001.0001.0001.000

Even with 492 directions and training it reaches 1.000 / 0.955 / 0.988 / 0.993 and is only approximately rotation-invariant, while Circ-PHT is exact with three directions and no training (Table 5 and Tables 10-12).

[1] Kaba et al., Equivariance with Learned Canonicalization Functions, ICML 2023. [2] Puny et al., Frame Averaging for Invariant and Equivariant Network Design, ICLR 2022. [3] Paik, Invariant Representations of Embedded Simplicial Complexes, 2023.

Would any result in Section 4 fail if the method were replaced by "rotate to the frame, then read the plain transform at any fixed set of directions"? What do the circles and the transform-based frame add?

Our inspiration for the usage of circles came from Figure 5 and the Motivating Example in Section 4. The proof of Theorem 4.4 and 4.11 did not need the structure of circles, but the structure of placing along circles was added as an inductive bias to better detect chirality (as reflections usually breaks the orientation along the circle, see Figure 5 and also Theorem 4.6).

We have now added Proposition 4.13 to justify our intuition, where we prove that the ordered circle method is strictly more expressive than the same directional measurements with their order forgotten (so as a set).

In Table 3 before the rebuttal, we also tested Circ-PHT to Circ-PHT with directions pooled. The latter does an averaging operation and forgets about the order and structure of circles, and performs worse than Circ-PHT.

We also empirically validated that preserving circular order does benefit learned read-outs. Using the same transform-guided frame, PHT landscapes, 12 directions, and 128-dimensional output, we compared circular convolution (readout is a circular CNN) with MLPs on labelled directions and on the flattened circular scan (which does not account the circular structure):

Read-outParametersBZRCOX2DHFR
Circular convolution127k0.7770.7510.818
MLP, labelled directions1.05M0.6890.7270.759
MLP, circular scan1.05M0.6620.6710.749

See Table 22 for more details. All three achieve 1.000 on clean retrieval, but the circular convolution is more robust with approximately eight times fewer parameters. The same ordering holds with only 10% of the training data (Table 22).

Is there a class of shapes and a fixed finite scan for which you can prove the descriptor is complete?

We added a new Proposition 4.14 that shows that the equivariant frame required by the theorem exists generically, so the completeness theorem applies generically. For any finite collection of pairwise inequivalent shapes satisfying its hypotheses, there exists a single finite scan, fixed for that collection, that distinguishes all its members.

For infinite classes, there need not be a set of fixed finite scans that can separate all the shapes apart (see Remark B.8 where we expplained why this is a well-documented challenge of the TDA community). In [4], the authors did show that, after placing lower bounds on the curvature of vertices, the number of scans needed to tell two shapes apart is finite. Their result does not establish a single fixed finite scan that distinguishes all shapes in the class. For point clouds with uniformly bounded cardinality, however, the directions can be fixed in advance.

[4] Curry, Mukherjee, Turner, How Many Directions Determine a Shape and other Sufficiency Results for Two Topological Transforms, 2022.

The experiments confirm the theorems but do not show what the method adds. In the retrieval test, the queries are exact rotated copies and the hard negatives are exact mirror images. For these inputs, Theorems 4.4 and 4.6 already guarantee a perfect score for any correct implementation [...]

Theorems 4.4 and 4.6 give invariance and describe how a reflection acts, but they do not say that a finite descriptor of a shape differs from that of its mirror image (Remark 4.7), and Theorem 4.11 gives separation in the limit, not at three directions. The observation that three directions are enough on four datasets is an empirical result (Table 2). The comparison with Paik (2023) above shows it is not automatic. Paik also proposed a reflection-sensitive pooling operator but establishes no completeness results. In our added experiments, even with 492 directions, the evaluated Paik variants do not achieve perfect retrieval across all four datasets (Appendix F.5, Tables 10–12), whereas Circ-PHT achieves this with three directions, no training, and a theoretical completeness result under explicit assumptions.

Theorems 4.4 and 4.6 establish rotation invariance and characterize how reflections act on the circular scan, while Theorem 4.11 provides a completeness guarantee under equidistributed refinement. In practice, Circ-PHT achieves perfect chirality-sensitive retrieval on all four datasets with just three directions (Table 2).

There is no comparison with a simple right-handed PCA alignment plus the same transform, or with the paper's frame plus a simple non-topological read-out. The frame ... can jump when the coordinates change a little. [...] How does Table 1 change with small coordinate noise, or with different conformers of one molecule [...]? If Circ-PHT holds up at least as well as the PCA pose, I would move to 6.

You were right that our transform-guided (TG) frame is not globally stable as its argmax changes the selected vertices. In Theorem 4.15, we gave a proof that the TG frame is locally stable, meaning that for geomtric graphs where the TG frame can be uniquely constructed, any sufficiently small perturbations of the graph admits a unique TG frame that varies continuously.

The inability to find an equivariant globally stable frame is a well-documented challenge in geometric and equivariant deep learning (see for example [5] & Prop 3.2 of [6]). This includes the PCA frame, and we explain in Remark B.17 why no equivariant frame can be globally stable. Thus, local stability is the best any researcher could hope for.

Our circle method is applicable to any frame function (Def 4.2), and the right-handed PCA frame you suggested would be an example. Thank you for the suggestion, and we added it to the paper as a second frame function (Appendix E.2.1). It is equivariant, so many of the theorems apply, and Circ-PHT with the PCA frame is exact without noise with three directions on the four datasets.

We empirically tested the TG and PCA frame and saw that PCA is the more stable frame, although it still showed signs of global instability (see Table 18 and 19). On these three datasets the PHT is more stable than point counting with either frame. The TG frame remains useful where PCA is not defined (repeated eigenvalues or vanishing third moments). The PCA frame also is insensitive of the topology of the graph, as it only depends on the vertices.

We ran the additional ablatons requested (Appendix G.1, G.2). For noise, we added Gaussian noise of standard deviation σ to every coordinate and repeat Table 1 with noisy queries. We also compared the method to a non-topological read-out by point-counting, where for each direction, we compute the fraction of vertices below each scan.

The following table displays the results on 3 directions, σ = 0.05 Å, mAP:

MethodBZRCOX2DHFR
Circ-PHT, transform-guided frame0.6620.6440.750
Point counting, transform-guided frame0.5980.5970.689
Circ-PHT, PCA frame0.9550.9830.968
Point counting, PCA frame0.9280.9270.963

For conformers we used 3DCS, which has several conformers per stereoisomer (Appendix G.4). Retrieving the other conformers of the same stereoisomer, Circ-PHT with the PCA frame reaches 0.826 mAP, against 0.424 with the transform-guided frame, 0.505 for MolAE and 0.276 for UniMol (Table 23). On the benchmark's own metric it reaches a zero-shot ES-AUC of 0.902 with four directions, up from 0.707 in the submission and above the best published value we know (0.782, MolAE).

[5] Dym et al. Equivariant frames and the impossibility of continuous canonicalization [6] Baker et al. An explicit frame construction for normalizing 3D point clouds.

It also does not say which shapes have such a frame at all.

The new Proposition 4.14 shows that the existence of such frames holds generically on compact non-empty subsets of R^n.

On 3DCS, how do versions that ignore reflections score, for example a 2-direction scan?

We added them (Table 24):

Descriptor3DCS ES-AUC
Circ-PHT, submission0.707
Circ-PHT, 2 directions (blind)0.710
Circ-PHT, order pooled (blind)0.820
Circ-PHT, PCA frame, 4 directions0.902

The blind versions score well because 3DCS is not a pure mirror-image benchmark: it contains diastereomers and conformers, and the conformers of two enantiomers are not mirror images of each other (closest pair at a median RMSD of 2.3 Å with a reflection allowed). Pooling the order also removes the dependence on the second frame axis, which was the unstable one. With the PCA frame the order-sensitive descriptor is the best in every split: 0.894 on enantiomer pairs and 0.905 on diastereomer pairs, against 0.796 and 0.838 for the order-pooled one (Table 25). We also found that the stored poses of 3DCS carry information: a PHT in laboratory coordinates gets 0.987 on the stored geometries and 0.517 after random rotations. Our descriptors are rotation-invariant and are not affected.

The reflected-copy experiment of Table 1 remains the direct test of chirality, and the new experiments show that the method is also robust to noise and conformers once the frame is stable. Thank you again for the suggestions, they made the paper much stronger.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

We thank the reviewers, the (senior) area, program, and general chairs for their valuable service for our community. The review process has helped us to increase the quality of our work. We have updated our manuscript, with most changes in blue, as well as the anonmyous repository.

Here we present a summary of the main points of the updates.

  • Robustness and frame construction. Per the suggestion of Reviewer BPn7, we added a right-handed PCA frame as an alternative to our original transform-guided (TG) frame (Definitions E.2–E.3) and studied both under coordinate perturbations (Appendix G.1, Tables 13–18). We also added a local stability result for the TG frame under explicit assumptions (Theorem 4.15), and we explained why local stability is the best any researcher could hope for (see Remark B.17). We also added experiments on real molecular conformers (Appendix G.4, Table 23). The PCA frame improves robustness considerably, including raising zero-shot ES-AUC on 3DCS from 0.707 to 0.902 (Appendix G.4, Table 24).

  • Understanding the role of circles and topology. We added Proposition 4.13, which shows that preserving the order of directional measurements is strictly more expressive than treating them as an unordered set. We added controlled experiments comparing circular sampling with fixed directions under the same PCA frame, directional transform, and sampling budget (Appendices G.2–G.3, Tables 19 and 21). We further investigated learned circular convolutions (Appendix G.3, Table 22) and other training-free read-outs (Appendix G.1, Table 14), to understand when the circular structure provides an advantage. We also compared the method to simpler Euler-curve and (non-topological) point-counting descriptors (Tables 19 and 21).

  • Completeness and theoretical assumptions. We clarified that equivariant frames exist generically (Proposition 4.14, proof in Appendix B.4), the scope of our finite-sampling guarantee (Theorem 4.11 and Remark B.8).

  • Broader comparisons and applicability. We expanded our comparison with Paik (2023) (Appendix F.5, Tables 10–12), investigated Hodge-Laplacian and spectral descriptors within the circle method (Appendix G.6, Tables 27–29), and added a synthetic experiment in four dimensions (Appendix G.8, Table 31). Finally, we measured computational costs on graphs with up to 10,240 vertices (Appendix G.7, Table 30).

  • Clarifying our contribution. We revised the main paper to clarify that other related works with our method. Our goal is to provide a general recipe for incorporating directional transforms into a framework that guarantees rotation invariance while retaining the ability to distinguish mirror images and thus detect chirality. In addition, our circle-based construction offers a useful inductive bias, with our experiments supporting its advantages over the other handcrafted approaches considered.

Thank you again for your service to the research community. Your feedback throughout the rebuttal process has helped us strengthen the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.