acceptodds
Under review as a conference paper at ICLR 2027

Heat Field Signatures: From Point Clouds to Smooth Geometry

Abstract

Bringing multiscale geometric analysis directly to irregular point clouds remains difficult: quantities such as local dimension, anisotropy, density variation, and geometric transitions are typically estimated through explicit neighborhood, manifold, or graph constructions, or left for neural networks to infer from coordinates. We introduce *Heat Field Signatures* (HFS), which lift a point cloud to a multiscale family of smooth ambient heat fields, providing a direct interface from discrete samples to geometric analysis. From this field, HFS computes closed-form global and local signatures directly from pairwise distances, capturing heat concentration, intrinsic dimension, anisotropy, and scale transitions. We further introduce the *Heat Dimension Spectrum* (HDS), a compact summary of multiscale geometric composition. HFS can be used as a closed-form descriptor, a lightweight learned representation, or a geometric feature channel for neural point-cloud models. Across synthetic and real-world benchmarks spanning subcellular, neuronal, tree, and protein data, HFS outperforms strong point-cloud and multiparameter-persistence baselines while substantially reducing end-to-end cost. On SCOP protein-fold classification, HFS improves over the strongest deep baseline by nearly 24 percentage points using coordinates alone, while standalone HFS representations are exactly rotation-invariant by construction. More broadly, HFS turns a classical heat field into a practical interface for multiscale geometric analysis in modern point-cloud learning.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
point clouds, geometric deep learning, representation learning, differential geometry, heat kernels, shape analysis, rotation invariance

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_HfpLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Heat Field Signatures (HFS), a multi-scale representation that maps irregular point clouds to smooth ambient heat fields and extracts geometric descriptors. The method combines closed-form global and local signatures with an optional learned pooling component. Experiments on synthetic and real-world point-cloud datasets show improved performance on several tasks and lower computational cost than some standard point-cloud neural-network baselines.

Critical strengths
  1. The proposed representation is relatively simple, interpretable, and avoids explicit graph or mesh construction while exposing interesting multi-scale geometric quantities.

  2. The empirical results are strong across several datasets, with additional advantages in computational efficiency and built-in rotation invariance.

Critical weaknesses
  1. A first concern is the degree of novelty and how the method is positioned relative to prior work. The ambient heat-field construction appears closely related to standard Gaussian smoothing and kernel-based density estimation. The main contribution therefore seems to lie in the geometric quantities derived from this field, their aggregation into the HFS/HDS representations, and their use for point-cloud learning. Although the paper discusses related methods such as HKS, diffusion maps, and intrinsic-dimension estimators, a more detailed comparison with prior work on Gaussian scale-space and kernel-based geometric analysis would help clarify which aspects of HFS are genuinely new.

There is also some missing literature in the section "Heat Kernels and Spectral Methods" such as prior work on defining spectral signatures on the graph Laplacian [1] and others using Fourier and wavelet analyses of the graph and Hodge Laplacians [2-4].

[1] NetLSD: Hearing the Shape of a Graph (KDD 2018, https://arxiv.org/abs/1805.10712)

[2] Graph Classification Gaussian Processes via Spectral Features (UAI 2023, https://arxiv.org/abs/2306.03770)

[3] Hodgelets: Localized Spectral Representations of Flows on Simplicial Complexes (ICASSP 2022, https://arxiv.org/abs/2109.08728)

[4] Graph and Simplicial Complex Prediction Gaussian Process via the Hodgelet Representations (NeurIPS 2026, https://arxiv.org/abs/2505.10877)

  1. The empirical evaluation is reasonably broad in terms of the number and diversity of datasets. However, many of the main benchmarks appear particularly well aligned with the type of multi-scale geometric structure that HFS is designed to capture. Evaluation on more generic point-cloud benchmarks, such as ModelNet40, ScanObjectNN, or ShapeNet, would therefore help assess the broader applicability and robustness of the method beyond these geometry-focused settings.

  2. The paper explicitly mentions the use of AI. The quality of Figure 2 is not satisfactory, especially for a conference such as ICLR. The figure is also overly crowded, with little explanation of its content.

Questions for authors
  1. Can Theorem 1 be extended to the data-dependent scale selection used by HFS in practice?

  2. Can a finite-sample guarantee analogous to Theorem 1 be obtained for the log-Hessian eigenvalues?

  3. How do the results change under normalisation-matched preprocessing and, for the synthetic datasets, matched point-cloud resolutions?

I have no other questions beyond those addressing my three concerns in the "Critical Weaknesses" section.

At the moment, I am leaning towards rejection. I would like to see these points addressed in order for me to consider increasing my score.

4 comments
AuthorsOpenReview ↗

Responses to Reviewer Hfpl - part 1

We thank the reviewer for the careful reading and constructive feedback. We address each point below; the General Response above summarizes the main revisions, including new baselines, matched-protocol experiments, a scalability study, and new theoretical results.

W1a: Novelty and relationship to Gaussian scale-space

A first concern is the degree of novelty and how the method is positioned relative to prior work.

We thank the reviewer for this thoughtful question, which goes to the heart of the paper. We agree that, up to normalization, uXu_X is a Gaussian kernel density estimate, and we claim no novelty for the field itself. The novelty is the role the field plays. We believe this may have been obscured by our extensive comparisons with deep 3D networks, which can suggest that HFS is another point-cloud classifier competing on their terms. Its aim is different.

The contribution: an interface between geometric analysis and machine learning. Learning has repeatedly advanced when a branch of mathematics gained a practical interface to finite data: kernels opened functional analysis, persistent homology opened topology, and graph Laplacians opened spectral geometry. Differential geometry has lacked such an interface for irregular point clouds. Intrinsic dimension, anisotropy, codimension, energy dissipation and scale transitions are classical, well-understood quantities, but a point cloud has no derivatives, so these tools could not act on it directly. HFS lifts the cloud to a smooth multiscale object on which these quantities become well defined, closed-form in pairwise distances and theoretically grounded.

How this differs from classical uses of Gaussian smoothing. KDE estimates a density at a selected bandwidth. Scale-space theory (Witkin, 1983; Lindeberg, 1994) tracks image structure on a grid. Hessian filters and ridge estimators (Frangi et al., 1998; Ozertem & Erdogmus, 2011; Genovese et al., 2014) detect one structure type at a chosen scale. HFS instead characterizes every sample at every scale through geometric invariants of the field:

  1. Scale as a measurement axis. Heat dimension, dheat=n+2t ∂tlog⁡uXd_{\mathrm{heat}} = n + 2t\,\partial_t \log u_X, is a scale derivative, not a spatial one. It recovers intrinsic dimension with a finite-sample guarantee (Theorem 1, now extended to the data-dependent scales used in practice, Appendix A.3), and its own scale derivative τ\tau localizes junctions, boundaries and mixed-dimensional transitions.
  2. Density-independent second-order geometry. The scale-normalized log-Hessian has universal eigenvalues, 00 in tangent and −1-1 in normal directions, so it gives a dimensionless tangent–normal decomposition at every point. It is a geometric-type descriptor, not a threshold-based detector, and needs no normal estimation.
  3. Density–geometry interaction. The logarithm separates density from geometry, and the Heat Dimension Spectrum measures how they interact (Δpk=pkρ−pkgeo\Delta p_k = p^\rho_k - p^{\mathrm{geo}}_k). Classical kernel methods do not represent this explicitly.

These readouts are not independent handcrafted features; the heat equation ties them together. For example, dheat=n+tr(2tQX)+2t ∥∇log⁡uX∥2,d_{\mathrm{heat}} = n + \mathrm{tr}(2tQ_X) + 2t\,\|\nabla\log u_X\|^2, so heat dimension counts the directions along which the structure extends, and the log-Hessian spectrum resolves which directions they are.

The geometry, not the smoothing, carries the signal. We tested this directly. A KDE-density baseline, which uses the heat field itself without its geometric readouts, trails the best method by 21.0 points on average (average rank 9.29). HFS-full trails by only 1.39 points (average rank 2.29, the best of all methods). Gaussian smoothing alone does not explain the results; the geometric analysis does.

Strong results where the geometry matters. HFS targets native point clouds, where the point set is the measurement and its heterogeneity carries the label: neuronal arbors, subcellular distributions, LiDAR returns, protein backbones and spatial point processes. There, HFS-full outperforms PTv2 and PointMamba on all four real datasets. On protein folds it more than doubles their accuracy (58.2% vs. 27.1% and 23.8%) using CαC_\alpha coordinates alone. On density-driven point patterns it reaches 97.6% (Processes) and 99.8% (DisksAnnuli) against multiparameter persistence, which models the same density–geometry interaction through topology. HFS approaches it through differential geometry.

Revision. We have added a paragraph, "Gaussian scale-space and kernel methods", to Section 2, added the KDE-density baseline to Table 1, and revised the introduction to state the geometric-interface perspective explicitly.

AuthorsOpenReview ↗

Responses to Reviewer Hfpl - part 2

W1b: Missing spectral literature.

There is also some missing literature in the section "Heat Kernels and Spectral Methods"

We thank the reviewer for these references, which we have added to the "Heat kernels and spectral methods" paragraph of Section 2. They also help clarify where HFS sits. Works [1–4] build heat-kernel, spectral, Fourier or wavelet signatures from the Laplacian of a given graph or simplicial complex. They diffuse on a combinatorial structure and analyze it through its spectrum. HFS addresses a different setting: unstructured point clouds, where no such structure is given. It analyzes the ambient field generated by the samples, without constructing a graph or complex and without any eigendecomposition. NetLSD [1] offers a useful parallel. Its heat-trace signature tr e−tL\mathrm{tr}\,e^{-tL} is a scale-indexed global heat summary of a graph, while our global energy E2(t)=1⊤P2t1E_2(t) = \mathbf 1^\top P_{2t}\mathbf 1, the sum of all entries of the ambient heat-kernel matrix, is its point-cloud analogue. HFS then goes further, adding local differential geometry (dimension, anisotropy and scale transitions) at every point.

W2: Generic benchmarks (ModelNet40)

The empirical evaluation is reasonably broad in terms of the number and diversity of datasets

We thank the reviewer for this suggestion. We have added ModelNet40 (Appendix D.4), alongside the two boundary datasets already in the paper. We would also like to clarify how we chose our main benchmarks, because it reflects the core of the paper. We are classifying point clouds, not the objects they were sampled from. The natural benchmarks for this task are ones where the point set itself carries the class: the density–topology benchmarks of the multiparameter-persistence literature, and real scientific data in which the measurement is a point cloud, spanning subcellular biology, neuroscience, forestry and structural biology. Object-recognition benchmarks such as ModelNet40, ScanObjectNN and ShapeNet test a different question: each cloud is a homogeneous sample of a surface, and the label is its global silhouette. Evaluating HFS on them is a bit like judging a tool built to read heterogeneity on data constructed to have none.

Results on ModelNet40. We use the official split, 1,024 points, identical unit-ball inputs for all methods and no augmentation, averaged over three seeds:

MethodOA (%)mAcc (%)
DGCNN92.3088.79
PointMamba88.7684.97
HFS-full86.1282.05
HFS+DGCNN92.0088.39

As expected, coordinate networks lead on the task they were designed for. HFS nevertheless remains competitive well outside its regime:

  • Standalone HFS-full reaches 86.1%, within 2.7 points of PointMamba, using only intrinsic geometric signatures and no coordinates.
  • HFS+DGCNN matches DGCNN to within 0.3 points, so adding HFS features does not degrade a strong backbone even on object recognition.

A clear division of labor. Together with the boundary datasets (IntrA and ATLAS-1), these results delineate the scope of HFS. On native point clouds, where geometry and density carry the label, HFS has the best overall performance and wide margins, such as more than doubling the accuracy of PTv2 and PointMamba on protein folds. On rigid-silhouette object recognition it is competitive but not dominant. A representation that computes intrinsic geometry should help exactly when that geometry carries the label, and this is what we observe.

Revision. Appendix D.4 now reports ModelNet40 together with the boundary datasets and discusses the scope of HFS.

W3: Pipeline figure

The quality of Figure 2 is not satisfactory

We thank the reviewer for this feedback. We first drew the pipeline figure in TikZ, but we were not satisfied with its visual quality, so we used Google Gemini to produce improved versions. Among the several candidates, the submitted PDF mistakenly included a near-final draft rather than the intended version. This use was disclosed in the AI Use Statement. In the revision, we have replaced Figure 2 with a cleaner version that has corrected labels and a simplified layout, and we now state the use of AI assistance directly in its caption. We will further simplify the figure and expand its caption for the final version, so that each stage of the pipeline is self-explanatory.

AuthorsOpenReview ↗

Responses to Reviewer Hfpl - part 3

Q1: Extending Theorem 1 to the data-dependent scale schedule.

Can Theorem 1 be extended to the data-dependent scale selection used by HFS in practice?

We thank the reviewer for this question. Yes: we have added this extension as Corollary 2 in Appendix A.3.

The key observation is that heat dimension varies smoothly with scale: its derivative in log⁡t\log t is the transition rate τ\tau, which is bounded. This lets us upgrade Theorem 1 from a single fixed scale to a guarantee that holds simultaneously over a whole window of scales and at every sample point, at the cost of only a logarithmic factor. Any scale chosen from the data, including our tmin⁡=cmin⁡rˉnn2t_{\min} = c_{\min}\bar r_{\mathrm{nn}}^2, then inherits the guarantee whenever it falls in this window.

Remark 3 in Appendix A.3 then shows that, as the number of points grows, the fraction of HFS's scheduled scales inside the valid window tends to one, with only the finest and coarsest scales remaining outside. This matches the design: the finest scales record sampling structure (clustering, repulsion and local density, which are the signal on point-process benchmarks such as Processes), the intermediate scales provably recover intrinsic dimension, and the coarsest scales record global organization.

Q2: Finite-sample guarantee for the log-Hessian eigenvalues.

Can a finite-sample guarantee analogous to Theorem 1 be obtained for the log-Hessian eigenvalues?

We thank the reviewer for this question. Yes. Intuitively, the scale-normalized log-Hessian at a point is a heat-weighted local covariance: it measures how far the neighbors spread in each direction, relative to the spread of the heat kernel itself. Along the structure, the neighbors spread exactly as the kernel does, which gives an eigenvalue near 00; across the structure they do not spread at all, which gives an eigenvalue near −1-1. The log-Hessian is therefore built from the same heat-weighted neighbor sums as heat dimension, only in matrix form, and the argument of Theorem 1 carries over with a matrix concentration inequality in place of the scalar one.

The two guarantees are closely linked: heat dimension is essentially the trace of the log-Hessian. Heat dimension counts the directions along which the structure extends, while the eigenvalues resolve which directions they are. We have added this extension, with its technical details, as Remark 2 in Appendix A.3.

Q3: Results under normalization-matched preprocessing and matched resolutions.

How do the results change under normalisation-matched preprocessing and, for the synthetic datasets, matched point-cloud resolutions?

We thank the reviewer for this question. In the revised Table 1, HFS receives exactly the same inputs as the deep baselines: centered, unit-ball clouds at identical point budgets, including resampled synthetic clouds. Appendix D.1 compares raw and matched inputs for all HFS variants. (We originally used raw coordinates at native resolution, the protocol of the multiparameter-persistence literature.)

Real benchmarks. Point budgets were always identical, so normalization is the only change. HFS-full:

AllenNeuroMorphoFOR-speciesSCOP
Raw97.974.266.660.5
Normalized97.974.064.158.2

The changes are at most 2.5 points, as expected: HFS sets its scales per cloud, so its geometric signatures are exactly invariant to translation and rescaling.

Synthetic benchmarks. HFS-full:

ProcessesOrbit5kDisksAnnuli
Raw, native resolution97.697.699.8
Normalized, matched resolution81.192.598.0

Orbit5k and DisksAnnuli change modestly. On Processes, native clouds range from 105 to 321 points because intensity differs across processes. Resampling to 256 points removes this information and adds duplicate points that corrupt the fine-scale pair statistics separating clustered, random and repulsive patterns. This is why multiparameter persistence uses native resolution, and we compare against it under that protocol (Table 4). Even so, every HFS variant still exceeds all deep baselines on Processes.

Overall. Under matched inputs, HFS-full has the best overall performance in Table 1 (average rank 2.29, average gap 1.39 points), ahead of PTv2 and PointMamba.

Revision. Table 1 uses matched inputs; the Table 4 caption states the shared raw protocol; Appendix D.1 compares raw and matched HFS on all seven benchmarks.

Dear Reviewer HfpL,

As the discussion period closes tomorrow, we wanted to check whether our responses and the revised paper address your concerns, and we would be glad to clarify anything before the deadline. Thank you again for your time and constructive review.

Authors

Reviewer_rWh7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Heat Field Signatures (HFS), which convert point clouds into multiscale smooth heat fields and extract geometric features such as density, dimension, and anisotropy. The method is validated on seven real-world benchmarks and shows strong performance compared with standard point-cloud methods.

Critical strengths
  1. Novel idea: HFS provides a simple way to explicitly capture multiscale geometry from irregular point clouds.

  2. Strong validation: The method performs well across several very different datasets, with useful ablation studies supporting the proposed components.

Critical weaknesses
  1. The comparisons are not fully matched: HFS and some baselines use different normalization or point-sampling settings, which may affect the reported performance gap.

  2. Scalability may be limited: The dense HFS computation has O(m2T)O(m^2T) complexity, which could be expensive for very large point clouds.

Questions for authors
  1. Could the authors provide a comparison where HFS and all baselines use exactly the same normalization and point budget?

  2. Could the authors show how runtime, memory, and accuracy scale when the number of points becomes much larger?

4 comments
AuthorsOpenReview ↗

Responses to Reviewer rWh7 - part 1

We thank the reviewer for the careful reading and constructive feedback. We address each point below; the General Response above summarizes the main revisions, including new baselines, matched-protocol experiments, a scalability study, and new theoretical results.

W1 & Q1: Matched normalization and point budgets

The comparisons are not fully matched.

We thank the reviewer for this point. In the revised Table 1, HFS receives exactly the same inputs as the deep baselines: centered, unit-ball clouds with identical point budgets and splits. Its advantage holds: HFS-full has the best overall performance (average rank 2.29, average gap 1.39 points), ahead of all baselines, including the newly added PTv2 and PointMamba.

Normalization barely affects HFS. On the real benchmarks, point budgets were always identical, so normalization is the only change:

AllenNeuroMorphoFOR-speciesSCOP
Raw97.974.266.660.5
Normalized97.974.064.158.2

The changes are at most 2.5 points. This is expected from the construction: HFS sets its diffusion scales per cloud, so its geometric signatures (heat dimension, log-Hessian eigenvalues, transition rate and HDS) are exactly invariant to translation and rescaling. Deep networks, in contrast, have fixed absolute length scales, such as ball-query radii and grid sizes, and require normalization [Qi et al., NeurIPS 2017; Qian et al., NeurIPS 2022].

Why the original submission used raw coordinates. Our two baseline families come with different native protocols. Deep networks normalize and resample; multiparameter-persistence methods, from whose literature our synthetic benchmarks originate, use raw coordinates at native resolution. We originally computed HFS once, following the latter, expecting normalization to matter little, which the experiment above confirms. We agree that the comparison with deep networks should use their protocol, and we now report each comparison under the protocol of its baselines: normalized inputs in Table 1, and raw native-resolution inputs in Table 4. The native protocol matters most for synthetic point patterns, where intensity within a shared observation window is part of the signal and resampling removes it. Under that protocol, HFS-full reaches 97.6% on Processes and 99.8% on DisksAnnuli.

Revision. Table 1 reports HFS on the deep networks' inputs; the Table 4 caption states the shared raw protocol; and the new Appendix D.1 compares raw and normalized HFS on all seven benchmarks.

AuthorsOpenReview ↗

Responses to Reviewer rWh7 - part 2

W2 & Q2: Scalability (runtime, memory and accuracy at large point counts)

Scalability may be limited.

We thank the reviewer for this question and have added a full scalability study (Appendix D.2). One clarification first: the dense cost is O(Tm2)O(Tm^2), quadratic in the number of points mm and linear in the number of scales TT, not O(mT2)O(mT^2). In practice, HFS scales better than standard point-cloud networks in memory and training cost, for a structural reason.

Linear memory, exact computation. HFS processes queries in fixed-size blocks, which reduces its memory from O(m2)O(m^2) to linear in mm without any approximation. Coordinate networks such as DGCNN build dense kk-NN graphs at every layer, which requires O(m2)O(m^2) memory. Because Gaussian contributions decay beyond radius O(t)O(\sqrt t), HFS further admits an O(mkT)O(mkT) truncated variant; all results below use the exact computation.

Training cost (Orbit5k, one H200, batch size 32):

PointsHFS-full memoryDGCNN memoryHFS-full stepDGCNN step
4,0960.78 GiB20.70 GiB6.9 ms166.7 ms
8,1921.49 GiB41.32 GiB12.8 ms442.9 ms
16,3842.91 GiBOOM22.1 msOOM
32,7685.75 GiBOOM40.6 msOOM
  • About 27×27\times less memory. DGCNN runs out of memory from 16,384 points on, while HFS trains at 32,768 points in under 6 GiB. Doubling the point count multiplies HFS memory by 1.94×1.94\times, versus 3.95×3.95\times for DGCNN, as linear versus quadratic storage predicts. The advantage holds at inference too: 1.1 GiB versus 12.1 GiB at 32,768 points.
  • Much cheaper training overall. HFS features are computed once and cached, and each step then runs on compact per-point features. Including the one-time extraction, the total training cost over a 250-epoch schedule is about 9×9\times lower than DGCNN's at 4,096 points and about 11×11\times lower at 8,192 points. Beyond that, DGCNN cannot be trained at all.

Accuracy at high resolution. Each method was retrained at every resolution on a fixed subset (400 train / 200 test clouds, identical unit-ball inputs, single seed; full-data results are in Table 1):

PointsOrbit5k: HFS-fullOrbit5k: DGCNNDisksAnnuli: HFS-fullDisksAnnuli: DGCNN
1,02492.093.582.551.5
4,09691.590.098.545.5
8,19292.092.598.046.5
16,38496.0OOM99.0OOM
32,76894.5OOM98.5OOM

HFS-full completes every resolution and stays accurate throughout. On DisksAnnuli, its accuracy improves with resolution, since denser sampling sharpens its density and dimension estimates. HFS therefore not only scales to large clouds but benefits from them.

Revision. We have added Appendix D.2, which covers complexity, runtime, memory, training cost and accuracy versus point count.

Reviewer_rWh7OpenReview ↗

Official Comment

I thank the author for addressing most of my concerns. I believe the contributions are sufficient for the venue, but do not stand out. I will maintain my score of 3 (weak accept).

We sincerely thank the reviewer for the constructive feedback and for recognizing our contribution. We truly appreciate the time and care you invested in reviewing our work.

Reviewer_khC6OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

To solve the issue that bringing multiscale geometric analysis directly to irregular point clouds, this paper proposes Heat Field Signatures (HFS), which lift a point cloud to a multiscale family of smooth ambient heat fields, providing a direct interface from discrete samples to geometric analysis.

Critical strengths

I think HFS is a simple and clear method without requiring a mesh, graph Laplacian, normal estimation. And authors provide some theoretical proofs in the Appendix to explain the proposed method. It helps me to understand the proposed method.

Critical weaknesses
  1. The paper claims their HFS "outperforms strong point-cloud baselines". However, in table1, the paper compares some "old" baseline like Point Transformer (2021), PointNet++ (2017), etc. The newest baseline in table 1 is published in 2021, which is really old because this paper is under review at ICLR 2027. The paper has to compare with some recent baselines, such as Point Transformer V2 (NeurIPS 2022), PointMamba (NeurIPS 2024), PointCHR (ICML 2026). It is necessary to evaluate whether the proposed method can outperform recent baselines instead of the "old" baselines to support the paper claim.

  2. The comparison protocol is not sufficiently controlled to convinced. As stated in Appendix C.1 (Lines 1122–1130), although all methods receive identical clouds and point budgets on the real benchmarks, the preprocessing differs. The baselines center each point cloud and scale it to the unit ball, whereas the HFS features are computed on raw coordinates. And the authors also point it. Therefore, these protocol differences make it difficult to determine whether the gains in table 1 arise from the proposed or from absolute-size information and differences in sampling resolution.

  3. From the definitions in Eq. (2) to Eq.(4) of Appendix A.2, Eq. (2) defines C2(t)=−t∂tlog⁡E2(t)C_2(t)=-t\partial_t\log E_2(t), while Eq. (3) defines E(t)=−12∂tE2(t)\mathcal{E}(t)=-\frac{1}{2}\partial_tE_2(t). Combining these two definitions shows that C2(t)=2tE(t)/E2(t)C_2(t)=2t\mathcal{E}(t)/E_2(t). Eq. (4) further defines E~(t)=E(t)/(E2(t)+ϵ)\widetilde{\mathcal{E}}(t)=\mathcal{E}(t)/(E_2(t)+\epsilon). Therefore, the four global quantities E2(t)E_2(t), C2(t)C_2(t), E(t)\mathcal{E}(t), and E~(t)\widetilde{\mathcal{E}}(t) are not mutually independent. This may weakens the claim in Appendix A.2 that “these quantities answer complementary questions”.

Questions for authors

The absence of recent baselines and the unfair comparison protocols make the results insufficiently convincing. Addressing these concerns would require extensive additional experiments, which means the paper needs to conduct major revision. Therefore, I tend to clear rejection. At last, the pipeline figure (Figure 2) seems like use the AI agent to produce it. Although this is not a primary reason for my recommendation rating, I encourage the authors to disclose this information directly in the figure caption or the main paper instead of only in the AI Use Statement.

4 comments
AuthorsOpenReview ↗

Responses to Reviewer khC6 - part 1

We thank the reviewer for the careful reading and constructive feedback. We address each point below; the General Response above summarizes the main revisions, including new baselines, matched-protocol experiments, a scalability study, and new theoretical results.

W1: Recent baselines

The paper claims their HFS "outperforms strong point-cloud baselines.

We thank the reviewer for this suggestion. We have added Point Transformer V2 (NeurIPS 2022) and PointMamba (NeurIPS 2024) to Table 1. We were unable to include PointCHR because its code is not publicly available. Both new baselines are trained from scratch on all seven benchmarks, on exactly the same inputs as every other method: centered, unit-ball clouds at identical point budgets and identical splits. The results strengthen our conclusion.

HFS has the best overall performance. Across all seven benchmarks:

MethodAvg. rank ↓Avg. gap to best ↓Worst-case gap ↓
PointMamba8.4317.7245.5
DGCNN6.2910.6127.7
PTv24.296.4531.0
HFS+DGCNN3.433.166.5
HFS-full2.291.394.1

HFS-full ranks first overall and is never more than 4.1 points from the best method on any benchmark. No baseline comes close to this consistency. Each modern backbone excels on some tasks and fails sharply on others: PTv2 leads on two synthetic benchmarks but trails by 31 points on SCOP, and PointMamba trails by 45 points on DisksAnnuli.

On real scientific data, the gap is clear. On identical inputs:

AllenNeuroMorphoFOR-speciesSCOP
PointMamba93.266.456.723.8
PTv296.773.062.627.1
HFS-full97.974.064.158.2

HFS-full outperforms both modern backbones on all four real datasets. On protein folds it more than doubles their accuracy, using CαC_\alpha coordinates alone.

Why newer architectures do not close the gap. The progress from PointNet to PTv2 and PointMamba consists of better ways to learn geometry from coordinates: attention, partition-based pooling, serialized state-space models. These advances target object-shape benchmarks such as ModelNet and ScanObjectNN, where a rigid silhouette is sampled densely and uniformly. Our data are different. Neuronal arbors, subcellular distributions, LiDAR returns and protein backbones are irregular point sets whose labels are carried by intrinsic multiscale geometry: local dimension, anisotropy, density variation and their interaction across scales. HFS computes these quantities explicitly, with a finite-sample guarantee, instead of asking a network to rediscover them from coordinates. The sparser and more irregular the data, the larger the advantage, as SCOP (128 points per cloud) shows most clearly.

On the synthetic Orbit5k and DisksAnnuli benchmarks, PTv2 leads HFS-full by 2.1 and 1.4 points under fixed point budgets. These are point patterns whose intensity is part of the data, and resampling to a fixed budget removes it. Under the native-resolution protocol used by multiparameter persistence, HFS-full reaches 97.6% on Processes and 99.8% on DisksAnnuli (Table 4, Appendix D.1).

Revision. Table 1 now includes PTv2 and PointMamba, together with average-rank and average-gap columns. Section 4 discusses the results, and we have revised the paper's claims to state the overall-performance result precisely.

AuthorsOpenReview ↗

Responses to Reviewer khC6 - part 2

W2: Comparison protocol

The comparison protocol is not sufficiently controlled to convinced.

We thank the reviewer for this careful observation. We have resolved this concern directly: in the revised Table 1, HFS uses exactly the same preprocessing as the deep networks, and its advantage remains.

Absolute size. On the real benchmarks, normalization is the only difference between the two protocols. HFS-full under raw and normalized inputs:

AllenNeuroMorphoFOR-speciesSCOP
Raw97.974.266.660.5
Normalized97.974.064.158.2

The changes are at most 2.5 points, and essentially zero on Allen and NeuroMorpho. On the same normalized inputs as all deep baselines, HFS-full has the best overall performance in Table 1 (average rank 2.29, average gap 1.39 points), ahead of PTv2 and PointMamba. Its gains therefore do not come from absolute-size information.

Sampling resolution. On the real benchmarks, all methods always received identical point budgets. On the synthetic benchmarks, HFS in Table 1 now uses the same resampled clouds as the deep baselines.

Why HFS is insensitive to normalization. HFS sets its diffusion scales per cloud, from the cloud's own nearest-neighbor distance and diameter. Its geometric signatures (heat dimension, log-Hessian eigenvalues, transition rate and Heat Dimension Spectrum) are therefore exactly invariant to translation and rescaling; only the raw density and energy amplitudes retain absolute scale. Deep networks, by contrast, have fixed absolute length scales, such as the ball-query radii of PointNet++ [Qi et al., NeurIPS 2017] and the grid sizes of PTv2 [Wu et al., NeurIPS 2022], and their optimization depends on coordinate magnitudes [Qian et al., NeurIPS 2022]. Normalization is a requirement of these architectures, not of HFS.

Why the original submission used raw coordinates. Our two baseline families come with different native protocols. Deep networks normalize and resample; multiparameter-persistence methods, from whose literature our synthetic benchmarks originate, use raw coordinates at native resolution. We originally computed HFS once, following the latter, expecting normalization to matter little, which the experiment above confirms. We agree that the comparison with deep networks should use their protocol, and we now report each comparison under the protocol of its baselines: normalized inputs in Table 1 and raw native-resolution inputs in Table 4.

Synthetic benchmarks. Here the protocol matters more, because the point pattern itself is the data: intensity within a shared observation window carries class information, and resampling to a fixed budget removes it. Under the native protocol shared with multiparameter persistence, HFS-full reaches 97.6% on Processes and 99.8% on DisksAnnuli (Table 4). Imposing the deep-network constraints on HFS is therefore the most conservative comparison we can make, and HFS still leads overall.

Revision.

  • Table 1 reports HFS on the same normalized inputs as the deep baselines.
  • The Table 4 caption and Section 4 state that HFS and the MP baselines share the raw, native-resolution protocol.
  • The new Appendix D.1 compares raw and normalized HFS across all seven benchmarks and analyzes the differences.
  • Appendix C.1 has been updated accordingly.
AuthorsOpenReview ↗

Responses to Reviewer khC6 - part 3

W3: Dependence among the global quantities.

From the definitions in Eq. (2) to Eq.(4) of Appendix A.2, Eq. (2) defines...

We thank the reviewer for this precise observation. The identity is correct and intentional: E=−12∂tE2\mathcal E = -\tfrac12\partial_t E_2 is the energy-dissipation identity of the heat equation, and it is exactly what makes the Dirichlet energy computable in closed form from pairwise distances. We agree that the word "complementary" should be made precise, and we have done so in the revision.

Functional dependence is not redundancy. Variance is an algebraic function of the first two moments, Var(X)=E[X2]−E[X]2\mathrm{Var}(X) = \mathbb E[X^2] - \mathbb E[X]^2, just as C2=2t E/E2C_2 = 2t\,\mathcal E/E_2; yet mean and variance are not redundant summaries. The same holds here, in three concrete ways:

  • Different geometric questions. E2E_2 measures how concentrated the heat is; E\mathcal E measures how fast that concentration dissipates; C2=−t ∂tlog⁡E2C_2 = -t\,\partial_t \log E_2 is the dimensionless log-rate, which approaches (n−d)/2(n-d)/2 in a dd-dimensional regime and therefore reads off codimension directly; and E~\widetilde{\mathcal E} is roughness per unit concentration, which retains the length scale at which dissipation occurs.
  • Different invariances. C2C_2 is dimensionless and invariant to rescaling the cloud, whereas E2E_2 and E\mathcal E scale with its size. They therefore carry different information to the classifier.
  • Not recoverable at finite scales. HFS samples T=8T=8 scales. E(ta)\mathcal E(t_a) is an exact derivative at each scale and cannot be reconstructed from the sampled values {E2(ta)}a=1T\{E_2(t_a)\}_{a=1}^T.

All HFS quantities are, by design, readouts of a single heat field, each corresponding to a classical quantity of geometric analysis: concentration, energy dissipation and codimension. This shared origin gives the representation its coherence: the readouts are linked by the heat equation rather than assembled as unrelated handcrafted features.

Revision. We have added a paragraph, "Relations among the global signatures", to Appendix A.2. It derives the energy-dissipation identity, states the relations C2=2t E/E2C_2 = 2t\,\mathcal E/E_2 and E~=C2/(2t)\widetilde{\mathcal E} = C_2/(2t) explicitly, and specifies the distinct role of each quantity.

AuthorsOpenReview ↗

Responses to Reviewer khC6 - part 4

Q1: Overall assessment and scope of revision.

The absence of recent baselines and the unfair comparison protocols make the results insufficiently convincing.

We thank the reviewer for the detailed review. We respectfully believe that the two concerns behind the rating, recent baselines and protocol control, are now fully resolved by experiments completed during the rebuttal period, and that they did not require a major revision. We also hope the reviewer will reconsider the contribution itself. Because our main comparisons are with deep 3D point-cloud networks, the paper may have given the impression that HFS competes with them on their own task, object recognition. Its goal is different.

What this paper contributes. Machine learning has advanced each time a branch of mathematics gained a practical interface to data: kernels opened functional analysis, persistent homology opened topology, and graph Laplacians opened spectral geometry. Differential geometry, one of the richest mathematical toolboxes, has lacked such an interface for irregular point clouds. Intrinsic dimension, anisotropy, codimension, energy dissipation and scale transitions are classical, well-understood quantities, but a point cloud has no derivatives, so they could not be computed on it directly. HFS fills this gap. By lifting a cloud to the multiscale heat field, it makes these quantities closed-form, theoretically grounded (Theorem 1 and its extensions in Appendix A.3) and exactly invariant, and so brings the machinery of geometric analysis into point-cloud learning. The geometry, not the smoothing, is what matters: a KDE baseline that uses the heat field without its geometric readouts trails the best method by 21.0 points on average, whereas HFS-full trails by only 1.39.

The setting. Most point-cloud benchmarks sample object surfaces homogeneously, so the label is a global silhouette, and deep networks are built for exactly that. HFS targets native point clouds, where the point set is the measurement itself and its heterogeneity carries the label: neuronal arbors, subcellular distributions, LiDAR returns, protein backbones and spatial point processes. Many important scientific problems produce data of this kind.

Both concerns are resolved, under conditions that favor the baselines.

  • Recent baselines. PTv2 (NeurIPS 2022) and PointMamba (NeurIPS 2024) are now in Table 1.
  • Protocol. HFS is now evaluated with exactly the deep networks' preprocessing: centering, unit-ball scaling and fixed-budget resampling. HFS does not need any of this, because its scales adapt to each cloud, and on synthetic point patterns the resampling removes information it can use. Table 1 is therefore the setting least favorable to HFS.
  • Even there, HFS-full has the best overall performance: an average rank of 2.29 and an average gap of 1.39 points, versus 4.29 and 6.45 for PTv2, the strongest baseline. It outperforms both modern backbones on all four real datasets, and on protein folds it more than doubles their accuracy (58.2% vs. 27.1% and 23.8%).
  • Normalization barely affects HFS on real data (at most 2.5 points), which rules out absolute size as the source of its gains (Appendix D.1). In its native setting, HFS-full reaches 97.6% on Processes and 99.8% on DisksAnnuli.

The revision also adds a scalability study, in which HFS has linear memory and trains at 32,768 points where DGCNN runs out of memory (Appendix D.2), along with new theory and an honest scope analysis on ModelNet40 (Appendix D.4). The method and its conclusions are unchanged; the new experiments strengthen them.

We believe HFS opens a genuinely new direction, bringing geometric analysis into machine learning through a single principled interface, and we would be grateful if the reviewer would reconsider the score in light of these results.

Figure 2. We thank the reviewer for raising this. We first drew the pipeline figure in TikZ, but we were not satisfied with its visual quality, so we used Google Gemini to produce improved versions. Among the several candidates generated, the submitted PDF mistakenly included a near-final draft rather than the intended version. This use was disclosed in the AI Use Statement. In the revision, we have replaced Figure 2 with the final version and now state the use of AI assistance directly in its caption.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful and constructive feedback. Before the individual responses, we briefly clarify what this paper contributes, since our comparisons with deep 3D networks may have suggested that HFS competes with them on their own task. It does not; it fills a different gap.

The contribution: bringing geometric analysis into machine learning. Learning has repeatedly advanced when a branch of mathematics gained a practical interface to data: kernels opened functional analysis, persistent homology opened topology, and graph Laplacians opened spectral geometry. Differential geometry has lacked such an interface for irregular point clouds. Intrinsic dimension, anisotropy, codimension, energy dissipation and scale transitions are classical tools, but a point cloud has no derivatives, so they could not be applied to it directly. HFS provides this interface. By lifting a cloud to the multiscale heat field, it makes these quantities closed-form, exactly invariant and theoretically grounded. The geometry, not the smoothing, is what matters: a KDE baseline that uses the field without its geometric readouts trails the best method by 20.56 points on average, versus 0.93 for HFS-full.

The task: classifying point clouds, not objects. Standard benchmarks such as ModelNet sample object surfaces homogeneously, so the label is a global silhouette, the setting deep networks are built for. HFS targets native point clouds, where the point set is itself the measurement and its heterogeneity carries the label: neuronal arbors, subcellular distributions, LiDAR returns, protein backbones and spatial point processes. Its natural comparison class is multiparameter persistence, which addresses the same density–geometry interaction through topology.

A conservative evaluation, and HFS still leads. At the reviewers' request, Table 1 now evaluates HFS under the deep networks' own protocol (centering, unit-ball scaling and fixed-budget resampling), which HFS does not need. This is the setting least favorable to it. Even there:

  • HFS-full has the best overall performance: an average rank of 2.00 and an average gap of 0.93 points, versus 4.00 and 6.00 for the strongest baseline, PTv2.
  • On real scientific data, it outperforms PTv2 and PointMamba on all four datasets, and on protein folds it more than doubles their accuracy (58.2% vs. 27.1% and 23.8%).
  • In its native setting, it reaches 97.6% on Processes and 99.8% on DisksAnnuli.

Main revisions.

  • Baselines: PTv2 (NeurIPS 2022), PointMamba (NeurIPS 2024) and a KDE-density baseline added to Table 1, with average-rank and average-gap columns.
  • Protocol: each comparison now uses the protocol of its baselines. Normalization changes HFS-full by at most 2.5 points on real data (Appendix D.1).
  • Scalability: linear memory. HFS trains at 32,768 points, where DGCNN runs out of memory (Appendix D.2).
  • Theory: Theorem 1 is extended to data-dependent scales and to the log-Hessian spectrum (Appendix A.3).
  • Scope: ModelNet40 and two boundary datasets, where coordinate networks lead as expected and HFS remains competitive (Appendix D.4).
  • Positioning: new related work on Gaussian scale-space, kernel methods and spectral signatures, and a revised introduction.
  • Code: we have added an anonymous code supplement containing implementations, experiment scripts, recorded results, and split definitions for Tables 1, 2, and 4 and the global-quantity ablation in Appendix D.5.

We believe HFS opens a new direction, a principled interface between geometric analysis and point-cloud learning, and we hope the reviewers will evaluate it in that light.

Dear Reviewers,

As the discussion period ends, we sincerely thank you for your time and engagement. Your comments led to substantial new work: modern baselines under a matched protocol, a full audit of all tables, an ablation of each global signature, a scalability study and new theory. Together, these have made HFS a clearer and stronger paper.

If our responses and revisions have addressed your concerns, we would be truly grateful if you would consider reflecting this in your final scores :)

Best wishes,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.