acceptodds
Under review as a conference paper at ICLR 2027

Persistence Positional Encoding: A Minimal Transformer Recipe for Persistence Diagrams

Abstract

Persistence diagrams (PDs) represent multiscale topological information as multisets of birth–death pairs . How best to use them with Transformers, however, remains unclear. Connecting PDs to Transformers requires choosing how topological information is tokenized and exposed to self-attention. Prior methods use specialized designs for tokenization, aggregation, Transformer blocks, or attention, but hard to compare since reported results rarely share PD inputs, splits, or model-selection rules. We instead confine the diagram-specific design to point representation, mapping each birth–death pair through fixed random Fourier features while leaving the Transformer backbone unchanged, and call this recipe Persistence Positional Encoding (PPE). In matched controls on eight TU graph benchmarks (TU8), changing only point representation shifts macro accuracy by 5.49 to 6.47 percentage points (pp), against 0.47 to 1.45 pp from changing the downstream module, identifying point representation as the dominant design variable tested. We further reevaluate vectorizations, kernels, learned representations, and Transformer-based models under one protocol with shared PD inputs, outer folds, and validation-only selection. Among all evaluated PD methods and downstream configurations, PPE achieves the highest mean on most benchmarks, including TU8, ORBIT5K, and MCB-C. Finally, we prove that grid averaging leaves a reconstruction error at any resolution, whereas a Transformer with fixed random Fourier frequencies is almost surely universal. PPE thus offers a minimal, Transformer-native baseline for persistence diagrams, and a common protocol for future comparison. Our code is available at the anonymous repository.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
transformers, representation learning, random Fourier features, topological data analysis, persistent homology, persistence diagrams

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_hhK4OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PPE. Each persistence-diagram point (b, d) is mapped through fixed random Fourier features, with frequencies sampled once from a Gaussian. The features are linearly projected, summed with a homology-dimension embedding, and processed by a standard Transformer with CLS readout. The authors compare Transformer-based diagram models, vectorizations, kernels and learned representations under a unified protocol. They report controls suggesting the point encoding matters more than the downstream module. They also prove a lower bound for grid reconstruction and almost-sure universal approximation with fixed frequencies.

Critical strengths
  • The evaluation is careful in a way that is rare in TDA/ML. All methods use identical diagrams, shared folds, validation-only selection and equal search budgets, and ablations separate the point map from the downstream network.
  • The method is simple, drop-in, and clearly outperforms the other Transformer-based diagram models (Table 1).
  • Theorem 2 makes explicit when random Fourier features are injective: at least three rows from a nondegenerate Gaussian, which holds almost surely.
Critical weaknesses
  • Limited novelty. PersLay and Hofer et al. already map each diagram point to a higher-dimensional vector before permutation-invariant aggregation. Transformers on diagrams also exist (Persformer, xPerT, MST). The new ingredient is a standard fixed Fourier map. The paper's own controls show the Transformer adds little on top of it: RFF+DeepSets reaches 72.20 TU8 vs 72.63 with the Transformer (Table 3).
  • Weak empirical case against non-Transformer baselines. The TU8 margin over ATOL+random forest is 0.66 pp (73.22 vs 72.56), with no spread or significance test. PPE has the best mean on only one of eight TU datasets (DHFR, Tables 1a and A3), and PersLay and persistence image+MLP beat it on ogbg-molhiv (Table A4). The claim of "highest mean on most benchmarks" therefore overstates the results.
  • The theory does not explain the results. Theorem 1 is the well-known fact that a fixed discretization cannot approximate arbitrarily well. It covers only hard binning (xPerT), not Persistence Images where smoothing is used, and does not touch the gains over grid-free models (Persformer, MST, point MLP). Theorem 2 rests on standard injectivity and moment arguments that could be cited rather than re-derived over two pages.
  • No discussion of stability. Stability is central to diagram vectorization and is the main guarantee of many cited baselines, yet it is never analyzed or discussed for PPE. This matters more because the paper emphasizes fine-scale coordinate differences. PPE gives near-diagonal points full-magnitude tokens.
Questions for authors
  • How sensitive is PPE to the frequency scale σ? It determines which scale of coordinate variation the encoder emphasizes. How should it be chosen relative to the range of the diagram coordinates, and how much does accuracy vary across the searched range?
  • Theorem 2 is stated for a topology in which diagonal points are retained and cardinalities must match. Does the result still hold for the bottleneck or Wasserstein topology, where near-diagonal points can be added or removed at small cost? If not, what does it imply about the stability of PPE?
  • Can you provide a stability result for PPE, or at least an empirical sensitivity analysis under small perturbations of the input data?
4 comments

We thank the reviewer for the detailed comments. Several insightful questions are addressed in the General Response (GC1–GC3); below, we address the review-specific points.

W1. About novelty.

Thank you for raising this concern. We agree that pointwise embeddings, Fourier features, and Transformers for PDs are all established. Our study is not aimed at achieving SOTA on a specific downstream task, but at understanding how PDs can be effectively represented for standard Transformers. As detailed in GC1, our novelty is not component-wise, but the controlled diagnosis and design principle. In particular, the 72.20 vs. 72.63 comparison is itself part of our evidence that self-attention is not necessary and that point representation is the dominant design variable tested. Please see GC1 for details.

W2. Comparison with non-Transformer baselines.

Thanks for pointing this out. We intended to state that PPE achieves the highest aggregate score on TU8 as a benchmark suite, and the highest mean on ORBIT5K and MCB-C. We understand that the current phrase “highest mean on most benchmarks” can be read more broadly, and we will revise the wording accordingly.

Role of non-Transformer baselines: Our goal is not to build a new SOTA model that outperforms all PD baselines. Instead, we ask how PDs should be represented so that a Transformer can use them effectively, as discussed in our response to W1. For this reason, we mainly compare PPE against Transformer-based models. Methods outside this scope, such as ATOL and PersLay, are included only as reference points. They help assess whether PPE makes effective use of the information in the PD. If PPE lagged far behind them, it would suggest that PPE fails to do so, even if it were the best among PD Transformer methods. In our results, PPE remains competitive with these references, suggesting that this is not the case.

The unified comparison is intended to place this minimal interface against established vectorizations and learned PD representations under the same protocol, rather than to claim uniform superiority over them. Accordingly, we report the full comparison under shared conditions, including cases where non-Transformer baselines perform better.

W3. Theory.

Theorem 1 is used to motivate PPE’s point-level tokenization, not to explain gains over PI or grid-free models. For Theorem 2, we already cite [1] and explicitly invoke Newton identities; however, these alone do not establish our theorem. We still need to show that universality holds when random frequencies are sampled once and then fixed. We will further examine this point as suggested. Please see GC2 for the intended scope of both theorems.

[1] Leshno et al., Neural Networks, 1993.

W4. Stability and near-diagonal points.

Thank you for the interesting question. For fixed frequencies, the Fourier encoder γ is Lipschitz: ‖γ(p )−γ(q)‖_∞ ≤ K‖p−q‖_∞ for some K. Thus, PD points matched under bottleneck/Wasserstein matching inherit a token-level stability bound. Combined with standard PH stability, this controls perturbations of matched PD points.

The remaining case is points matched to the diagonal. PPE retains low-persistence points rather than suppressing them a priori, since short-lived features can still carry meaningful geometric information [2,3]. In our plain-vs-sprinkle donut task (Sec. 4.4; Table A5), where the classes differ only by sub-cell coordinate shifts, PPE reaches 71.97% versus 54.11/52.32/50.91% for xPerT/Persformer/MST. Additionally, The empirical sensitivity analysis under raw-input perturbations is provided in GC3.

[2] Feng & Porter, SIAM Review, 2021.

[3] Stolz et al., Chaos, 2017.

Q1. Sensitivity to the frequency scale σ.

We varied only σ. Values are mean ± population SD over three seed-wise 10-fold means (seeds 101–103; 30 fits/σ).

σTU8 (8 graph datasets) (%)ORBIT5K (%)
168.98 ± 0.3380.19 ± 8.90
270.47 ± 0.2084.60 ± 3.37
471.72 ± 0.1389.85 ± 0.17
872.24 ± 0.3490.57 ± 0.18
1671.84 ± 0.4290.67 ± 0.15
3272.13 ± 0.3790.65 ± 0.24

Very small σ reduces performance, while results are stable over σ ∈ {8,16,32}, varying only 0.40 pp on TU8 and 0.10 pp on ORBIT5K. We found no consistent relation between birth–death coordinate range and preferred σ. Thus, our main experiments select σ by validation-only random search, not from the PD coordinate range.

Q2. Does Theorem 2 hold under the bottleneck/Wasserstein topology?

Yes. Theorem 2 uses a finer topology than the standard bottleneck/Wasserstein topology (τ_std ⊂ τ_fine). Hence, every τ_std-continuous function is also τ_fine-continuous, i.e., C(τ_std) ⊂ C(τ_fine). Therefore, Theorem 2 already covers all target functions continuous under the standard topology.

Q3. Empirical sensitivity analysis.

Thank you for the suggestion. Due to space constraints, please refer to GC3 for our response to this point.

I thank the authors for their detailed response and additional experiments. I appreciate the careful methodology, which allows the contributions of point representation and downstream architecture to be disentangled, and I find the clarified claims appropriately supported. Some questions about stability remain, particularly since pointwise Lipschitz continuity does not establish stability of the complete diagram representation. Overall, while I consider the experimental study solid and valuable, my reservations about the limited methodological and theoretical novelty remain, and I therefore maintain my original rating.

AuthorsOpenReview ↗

Response (1/2)

Thank you for the continued discussion. We appreciate that the reviewer now finds the controlled methodology valuable and the clarified claims appropriately supported. We nevertheless would like to briefly clarify why we view the contribution as methodologically meaningful despite using established components.

On methodological and theoretical novelty.

While the individual components of PPE are established, the novelty of individual components and that of the overall methodology need not coincide. A relevant precedent is the original Vision Transformer (ViT; Dosovitskiy et al., ICLR 2021), whose central architectural idea was also simple—tokenizing images into patches and applying an essentially standard Transformer. Its significance lay not in inventing each component, but in establishing an effective interface between a new data modality and a general-purpose architecture. We do not claim that PPE has the same scope or impact as ViT; the analogy is only about this distinction.

Our contribution is similarly not “RFF + Transformer” itself. The controlled finding is that, among the choices tested, point representation is substantially more consequential than downstream architectural specialization: changing the point encoding yields 5.49–6.47 pp under matched controls, whereas changing the downstream module yields only 0.47–1.45 pp. This identifies a previously unestablished design principle for PD learning and shows that a standard Transformer can use PDs effectively without PD-specific attention or pooling.

The theoretical contribution should be interpreted at the same level. Theorem 1 establishes a decoder-independent limitation of uniform grid averaging, while Theorem 2 establishes that frequencies sampled once and then fixed retain universal approximation in our setting. To our knowledge, these precise results have not previously been established in this setting. We do not claim that these results are intended as deep mathematical contributions in isolation. Rather, we believe that, in an AI paper, a theoretical result can be meaningful when it establishes a previously unproven property of the proposed methodology and directly informs its design. In this sense, the two theorems provide new, targeted theoretical support for PPE’s deliberate representation choices.

Together with the unified evaluation protocol and the strong improvements over prior PD-Transformer methods, we believe the contribution is therefore more than an application of an existing Fourier map: it provides a controlled methodological finding, a minimal PD-to-Transformer interface, and theoretical justification for its specific design choices.

AuthorsOpenReview ↗

Response (2/2)

Regarding the reviewer’s remaining stability concern, our previous response established Lipschitz stability for matched PPE tokens and discussed diagonal matching. Here, we make explicit how this pointwise result extends to the complete PPE output and, together with standard PH stability, to raw-input perturbations.

Sensitivity analysis.

GC3 directly addressed the reviewer’s request for “or at least an empirical sensitivity analysis under small perturbations of the input data.” PPE remains robust under small raw perturbations and especially direct PD perturbations: at ε=10−4ε=10^{-4} and 10−310^{-3}, only 0.17% and 1.90% of predictions change, versus 10.27% and 47.80% for Persistence Images (PI). Thus, GC3 shows that the trained PPE model is empirically robust to small perturbations.

For the standard bottleneck topology, let ε=dB(D,D′)ε=d_B(D,D'). A bottleneck matching has two cases: (i) off-diagonal points matched to off-diagonal points, where every feature with persistence >2ε>2ε must be matched off-diagonal; and (ii) points matched to the diagonal, which necessarily have persistence ≤2ε≤2ε.

(i) From pointwise to complete matched-diagram stability (persistent off-diagonal features).

Let D,D′D,D' contain nn matched points with label-preserving bijection ππ satisfying

max⁡x∈D‖x−π(x)‖∞≤ε.\max_{x∈D}‖x-π(x)‖_∞ ≤ ε.

For

K=2π(∑j‖ωj‖12)1/2,K=2π(∑_j‖ω_j‖_1^2)^{1/2},

for RFF frequencies ωjω_j, γΩγ_Ω satisfies

‖γΩ(x)−γΩ(π(x))‖2≤Kε.‖γ_Ω(x)-γ_Ω(π(x))‖_2 ≤ Kε.

Since

zk(x)=WinγΩ(x)+ek,z_k(x)=W_{in}γ_Ω(x)+e_k,

with input projection WinW_{in} and label embedding eke_k, which cancels for matched labels; for token matrix ZDZ_D (Frobenius norm) and induced permutation PπP_π,

‖ZD−PπZD′‖≤n1/2‖Win‖opKε.‖Z_D-P_π Z_{D'}‖ ≤ n^{1/2}‖W_{in}‖_{op}Kε.

PPE has no sequence positional encoding, so ππ does not affect the CLS output. The fixed Transformer/readout is Lipschitz on this bounded fixed-cardinality domain, with finite nn-point constant Λθ,nΛ_{θ,n},

‖Fθ(D)−Fθ(D′)‖≤Λθ,n‖Win‖opKn1/2ε,‖F_θ(D)-F_θ(D')‖ ≤ Λ_{θ,n}‖W_{in}‖_{op}Kn^{1/2}ε,

where Fθ(D)F_θ(D) is the PPE logit vector.

(ii) Low-persistence features matched to the diagonal.

If ε=dB(D,D′)ε=d_B(D,D'), every point xx matched to the diagonal ΔΔ satisfies

d∞(x,Δ)≤ε,d_∞(x,Δ)≤ε,

hence has persistence at most 2ε2ε. Thus, diagonal matching is confined to low-persistence features near the diagonal.

PPE does not suppress such points at the tokenizer; instead, the downstream Transformer can learn whether near-diagonal features are task-irrelevant topological noise or informative signal.

For diagrams with at most MM off-diagonal points, suppose uniformly, for some CΔC_Δ, that the trained model satisfies

‖Fθ(D⊔{x})−Fθ(D)‖≤CΔd∞(x,Δ).‖F_θ(D⊔\{x\})-F_θ(D)‖ ≤ C_Δ d_∞(x,Δ).

Define

Lθ,M:=max⁡1≤n≤MΛθ,n<∞.L_{θ,M}:=\max_{1≤n≤M}Λ_{θ,n}<∞.

Under this condition,

‖Fθ(D)−Fθ(D′)‖≤(Lθ,M‖Win‖opKM1/2+2MCΔ)dB(D,D′).‖F_θ(D)-F_θ(D')‖ ≤ (L_{θ,M}‖W_{in}‖_{op}KM^{1/2}+2MC_Δ)d_B(D,D').

Indeed, off-diagonal matched pairs are controlled by (i), while each diagonal-matched point contributes at most CΔεC_Δε. Since each diagram has at most MM off-diagonal points, their total contribution is at most 2MCΔε2MC_Δε.

Combined with standard PH stability for the chosen filtration,

dB(D(X),D(X′))≤CPHdraw(X,X′),d_B(D(X),D(X')) ≤ C_{PH}d_{raw}(X,X'),

where drawd_{raw} is the raw metric and CPHC_{PH} the PH-stability constant, this yields end-to-end stability from raw perturbations to PPE logits under the same near-diagonal condition, providing a conditional counterpart to GC3’s raw-data perturbation experiment.

Why not suppress near-diagonal points a priori?

This is a deliberate modeling choice. Short-persistence features are often treated as topological noise, but long-persistence features alone do not universally determine downstream relevance. Prior work shows that short-lived topological features can encode meaningful geometric or data-dependent structure [1–3]. Therefore, forcing every near-diagonal point to vanish before learning can discard information useful for a downstream task.

Our plain-vs.-sprinkle donut experiment is consistent with this motivation. Its discriminative signal depends on fine-scale PD-coordinate differences, and PPE outperforms the strongest competing PD-Transformer method by 17.8617.86 pp. This supports preserving fine-scale point-level information rather than suppressing features solely by persistence: PPE exposes individual points to the downstream Transformer and learns their task-dependent relevance.

[1] Michelle Feng and Mason A. Porter. Persistent Homology of Geospatial Data: A Case Study with Voting. SIAM Review, 63(1):67–99, 2021.

[2] Peter Bubenik, Michael Hull, Dhruv Patel, and Benjamin Whittle. Persistent Homology Detects Curvature. Inverse Problems, 36(2):025008, 2020.

[3] Bernadette J. Stolz, Heather A. Harrington, and Mason A. Porter. Persistent Homology of Time-Dependent Functional Networks Constructed from Coupled Time Series. Chaos, 27(4):047410, 2017.

Reviewer_r6X9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper discusses how to apply Persistent Diagrams (DP) to transformers. The proposed method is based on Fourier analysis and is designed to preserve the point-based representation of DP. By theoretically demonstrating the limitations of grid-based approaches, the paper highlights the advantages of the proposed point-based method and confirms its superiority over conventional methods through experiments.

Critical strengths

While many PD learning methods involve data degradation, the Transformer approach—which minimizes information loss using point-based criteria—is intriguing. It is also significant that it has yielded better experimental results than other Transformer-based PD methods.

Critical weaknesses

The concern regarding this paper is whether it is appropriate in terms of both content and format for an AI conference. Although it focuses on the “Transformer” learning method, the discussion in this paper is confined to the scope of PD. This fact raises doubts as to whether this discussion is appropriate for an AI conference. Discussions in the field of AI should be grounded in the subject matter and challenges being addressed; narrowing the scope to a specific method risks being arbitrary. To put it bluntly, if a method that does not use PD outperforms the PD-based Transformer method on the benchmark dataset, there is no value in discussing the PD-based Transformer method. If the focus is solely on how PD is handled, the discussion should be considered part of the TDA (Transformer-based Deep Learning) field. From this perspective, the same applies to xPerT, PersFormer, and MST—which were used as comparators—and this likely explains why they were not accepted as papers. On the other hand, methods such as PersLay were also compared with approaches outside the TDA framework. From this standpoint, there are significant concerns regarding whether this paper is suitable for an AI conference.

Questions for authors

Given the same problem formulation, is it possible to present a comparison of the results—as well as the relationships—between the proposed method, state-of-the-art methods that do not use PD, and transformer-based methods in a way that demonstrates the value of discussing the proposed method?

6 comments

We appreciate the reviewer’s comments. To clarify the scope of our work, we kindly ask the reviewer to first read the General Response (GC1–GC3). Below, we respond to the review-specific points.

W1 & Q1. Scope, Relevance, and Conference Fit

Thank you for raising these concerns. We agree that broader comparisons help position the work. However, higher task accuracy from a non-PD method does not by itself invalidate a contribution to PD representation learning. Our question is not which model achieves the highest accuracy on graph or point-cloud benchmarks, but how PDs should be represented so that a standard Transformer can use them effectively. Transformers have become general-purpose backbones across a wide range of modalities and learning problems. Therefore, how such explicit topological information can be effectively exposed to a standard Transformer is itself a broadly relevant representation-learning problem. A reusable PD-to-Transformer interface could make explicit topological information easier to incorporate into Transformer-based pipelines.

PiPE (ICML 2025) establishes complementary graph-distinguishing capabilities of persistent homology and common positional encodings, and shows expressivity gains from combining them [1]. This shows that persistent homology can complement other input representations rather than compete with them. FILTR (CVPR 2026) probes several frozen pretrained 3D Transformer encoders on the DONUT topology benchmark [2]. Even the best-performing representation, the Point-BERT CLS token, achieves only 57.2% accuracy for connected-component prediction and 25.9% for genus prediction. These studies suggest that strong pretrained representations do not necessarily capture topological information effectively on their own.

Our goal is to enable Transformers to effectively capture topological structure. In fact, PDs provide a well-motivated representation for this purpose. PD are a standard explicit representation of multiscale topological information produced by persistent homology, which has been used across graphs, point clouds, molecular data, and other scientific domains. Persistent-homology-based learning has also been actively studied in major ML venues, including NeurIPS, ICLR, ICML, and AISTATS.

Our experiments identify a concrete representation bottleneck under a shared protocol. Prior Transformer-based PD methods have struggled to effectively utilize the information contained in PDs. In Table 2, even simple conventional vectorizations—persistence images, persistence landscapes, and Betti curves with random forests—reach TU8 accuracies of 71.67%, 71.17%, and 71.01%, respectively, while xPerT, PersFormer, and MST achieve only 69.32%, 66.06%, and 64.68%. However, PPE reaches 73.22% with a standard Transformer. More importantly, Table 3 isolates the design choice: replacing a point MLP with RFF under the same Transformer improves TU8 from 67.14% to 72.63%, whereas changing RFF+DeepSets to RFF+Transformer yields only 0.43 pp. Thus, point representation is more consequential than backbone choice in the settings tested.

Regarding Q1, we distinguish task-level references from controlled representation comparisons. Table A4 in Appendix A already includes non-PD GIN, GCN, and PointNet-style references. On ogbg-molhiv, GIN reaches 73.10 ROC-AUC versus PPE’s 69.32. On MCB-C, PPE reaches 65.79 balanced accuracy versus 63.76 for the PointNet-style reference. These results provide broader task-level context. We do not claim that PPE is consistently state-of-the-art for these downstream tasks. Rather, our results show that a standard Transformer operating only on explicit topological information can reach competitive task-level performance relative to raw-input models.

Our contribution is therefore a minimal PD-to-Transformer interface and controlled evidence about the role of representation, not a claim that PDs should replace raw-input methods. This positioning is consistent with ICLR’s stated scope, which includes learning on graphs, geometries, and topologies [3].

[1] Verma, Y., Souza, A. H., and Garg, V. K. Positional Encoding meets Persistent Homology on Graphs. ICML, 2025.

[2] Martinez, L., and Ovsjanikov, M. FILTR: Extracting Topological Features from Pretrained 3D Models. CVPR, 2026.

[3] ICLR 2027 Call for Papers.

Thank you to the authors for their detailed and thoughtful response.

I agree that the proposed method is not required to outperform all other existing approaches. However, I remain skeptical about whether the core objective of the paper—"how PDs should be represented so that a standard Transformer can use them effectively"—is inherently suitable as a central theme for an AI conference. Since the goal is explicitly centered on utilizing PDs, the discussion naturally leans toward a TDA perspective, and it is not immediately clear why Transformers have an inherent necessity to leverage PDs.

Therefore, it is essential to clarify the unique value gained by incorporating PDs into Transformers that cannot be achieved through other methods. While comparing the proposed method against others is a straightforward approach, another valid direction would be to demonstrate that PDs provide information that is uniquely useful for the specific objectives of Transformers and cannot be obtained via alternative representations. While I acknowledge that learning based on PDs has been discussed at major AI conferences, my understanding is that such papers are generally accepted only when they effectively incorporate the aforementioned perspective.

Given these considerations, I will maintain my current score.

AuthorsOpenReview ↗

Response (1/2)

We thank the reviewer for the continued discussion. While we respect the reviewer’s assessment, we would like to further discuss the relevance of our work to representation learning, as we believe there is an important distinction between establishing the necessity of a particular representation and studying how that representation can be effectively learned and utilized.

1. Representation learning and conference relevance.

The reviewer asks why Transformers have an inherent necessity to leverage PDs, and whether PDs provide information that cannot be obtained through alternative representations. However, we do not claim that Transformers must use PDs, or that PDs contain information fundamentally unavailable from every alternative representation. Since PDs are themselves derived from the original input, a sufficiently expressive model may in principle recover similar information.

Rather, the value of a representation often lies not in creating new information, but in making existing information more accessible to a learning model. PDs explicitly expose multiscale topological structures that are otherwise implicit in complex inputs such as graphs and point clouds. Such structured representations can provide useful inductive biases, potentially making relevant information easier to learn with finite data and model capacity.

Importantly, this complementary value is not merely hypothetical. As noted in our previous response, PiPE (ICML 2025) formally establishes graph-distinguishing capabilities of persistent homology that are complementary to conventional positional encodings, while FILTR (CVPR 2026) shows empirically that strong pretrained 3D representations can still fail to capture basic topological properties. Thus, our motivation is not that Transformers inherently need PDs, but that PDs provide an established structured representation that explicitly exposes complementary topological information.

We believe this is precisely one of the fundamental motivations of representation learning: organizing and exposing information already present in the data in a form that general-purpose learning architectures can exploit effectively. Our work addresses this question by investigating how explicit topological information should be represented for a standard Transformer, and identifying which design choices matter most through controlled experiments. We therefore believe this is a legitimate representation-learning problem well aligned with ICLR, rather than a question confined to TDA.

2. A relevant precedent: Vision Transformer.

A relevant example is the original Vision Transformer (ViT; Dosovitskiy et al., ICLR 2021). Its central architectural idea was remarkably simple: tokenize images into patches and process them using an essentially standard Transformer. Interestingly, the original ViT even underperformed comparable CNNs when trained on ImageNet-scale data, requiring substantially larger pretraining datasets to become competitive.

Under an evaluation criterion that requires a representation to be indispensable to its backbone or to consistently outperform alternative architectures, one could similarly question the methodological significance of the original ViT. Nevertheless, its importance lies in demonstrating that general-purpose Transformers can effectively process images without CNN-specific architectures, establishing an important direction for subsequent research.

We certainly do not claim that PPE has the same scope or impact as ViT. Rather, this example illustrates that studying how an established data modality can be effectively interfaced with a general-purpose architecture is itself a meaningful methodological question, even when alternative representations and architectures exist.

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

The paper proposes encoding individual persistence-diagram points using fixed random Fourier features, together with a learnable homology-dimension embedding, before processing them using a standard Transformer. Its main empirical finding is that point representation contributes substantially more than the downstream architecture under the tested configurations. The authors support this through controlled experiments, a unified benchmark evaluation, and two expressivity results.

Critical strengths

The main strength of this paper is that it addresses a fundamental question in learning from persistence diagrams: how much of the representational power comes from the encoding of individual topological features, and how much comes from the architecture used to process them?

The experimental analysis provides useful evidence that the choice of representation can be more consequential than the downstream architecture. I find this observation interesting, as it has implications beyond the particular model proposed.

Another strength is the simplicity of the construction and the systematic comparison with existing approaches. The authors show that a standard Transformer can effectively process persistence diagrams without introducing specialized attention mechanisms or aggregation operators. The unified experimental protocol is also valuable, particularly in a field where differences in filtrations, representations, and evaluation procedures make comparisons difficult.

Critical weaknesses

My main concern is the connection between the theoretical results and the actual contribution of the proposed representation. The universal approximation result establishes expressivity under a particular topology on bounded multisets, but it is not clear what this tells us about learning stable functions on persistence diagrams. In particular, the result does not seem to distinguish the proposed construction from other sufficiently expressive pointwise representations. Similarly, the grid reconstruction result establishes information loss for a particular form of discretization, but does not necessarily imply a fundamental limitation of grid-based approaches in learning functions of persistence diagrams.

A second concern is the role of the Transformer itself. The experimental results suggest that most of the improvement comes from the Fourier representation, while the differences between downstream architectures are relatively small. This raises the question of whether the proposed method should primarily be understood as a representation technique for persistence diagrams rather than a Transformer-specific contribution. I think the paper could be more precise about this distinction.

Questions for authors

--What is the fundamental property of the Fourier representation that makes it particularly suitable for persistence diagrams? Is the improvement related to its frequency content, its ability to distinguish points, or some property specific to persistent homology?

--How does the proposed construction relate to stability? In particular, can the authors characterize how small perturbations of a persistence diagram, measured using standard diagram distances, affect the resulting representation and prediction?

--The universal approximation theorem uses a topology that requires eventual agreement in diagram cardinality. How does this relate to the usual topology on persistence diagrams, where points may appear or disappear near the diagonal? Can the result be extended to this setting?

Given that the improvements from the Transformer over simpler permutation-invariant architectures appear relatively small, what aspects of the learning problem actually require self-attention? Would the main conclusions remain unchanged if the Transformer were replaced by a sufficiently expressive DeepSets architecture?

my assessment : I find the paper interesting, particularly its emphasis on representation rather than architectural complexity. The experiments are systematic and the proposed method is simple and effective. My main reservation concerns the interpretation of the theoretical results and whether they establish anything fundamental about the proposed construction beyond the expressive power of Fourier features.

I lean toward acceptance because the empirical analysis offers a useful contribution, although I believe the theoretical claims should be interpreted more carefully.

1 comment

We thank the reviewer for the positive and constructive assessment. Several insightful questions are addressed in the General Response (GC1–GC3); below, we address the review-specific points.

W1 & Q2 & Q3. Scope of the theoretical results and stability

W1: "My main concern is the connection between ..."

Q2: "How does the proposed construction ..."

Q3: "The universal approximation theorem uses ..."

These questions about the theorem, topology, and stability are closely related, so we address them together here. The theorems are intended to provide additional justification for specific design choices in our model. Due to space constraints, please see GC2 for the intended scope of the theorems. We will revise the corresponding statements to clarify our intended scope. For the topology-related discussion of Theorem 2, please see hhK4-Q2; for stability, please see hhK4-W4; and for the perturbation-based empirical sensitivity analysis, please see GC3.

W2 & Q4. Role of the Transformer.

W2: "A second concern is the role of the Transformer ..."

Q4: "Given that the improvements from the Transformer ..."

Thank you for raising this point. As discussed in GC1, our goal is not to show that self-attention is necessary, but to study how PDs can be represented for a standard Transformer. Table 3 shows that the core gain comes from representation: with RFF fixed, DeepSets and Transformer reach 72.20%→72.63% (+0.43 pp), whereas replacing raw coordinates with RFF under DeepSets improves 64.92%→72.20% (+7.28 pp). Thus, yes, our main representation-level conclusion remains unchanged with DeepSets. The core representation principle is not Transformer-specific. However, the Transformer is our primary setting because we study the PD-to-Transformer interface.

Q1. Why Fourier features for persistence diagrams?

Q1: "What is the fundamental property ..."

PDs encode information through the spatial configuration of birth–death points, and task-relevant structure can occur at different spatial scales on this plane. We therefore believe that the main advantage of the Fourier representation is its frequency content, rather than injectivity alone or a property unique to persistent homology.

A persistence point x=(b,d)∈ℝ² is mapped by PPE to γ(x) = [cos(2πΩx); sin(2πΩx)]. Here, frequency refers to spatial frequency on the birth–death plane. For two nearby points x and x + δ, (1/m) ‖γ(x)−γ(x+δ)‖₂² = (4/m) Σⱼ sin²(πωⱼᵀδ),

and, for sufficiently small δ, (1/m) ‖γ(x)−γ(x+δ)‖₂²≈(4π²/m)Σⱼ(ωⱼᵀδ)².

Thus, the response of each Fourier component to a coordinate displacement is governed by ωⱼᵀδ. Higher-frequency components have shorter spatial wavelengths and provide basis functions for representing variation at finer spatial scales on the birth–death plane. This does not conflict with persistence stability: stability bounds the magnitude of diagram changes under small input perturbations, but does not imply that all small, systematic coordinate variations are task-irrelevant.

Importantly, injectivity alone does not explain the gain: raw coordinates are already injective, and a sufficiently expressive MLP can in principle represent the same functions. The practical distinction is optimization and inductive bias. A raw-coordinate MLP must learn rapidly varying functions of (b, d) through optimization, whereas Fourier features explicitly expose basis functions at multiple spatial frequencies. The downstream network can therefore select and combine these components rather than having to construct the corresponding high-frequency structure from raw coordinates. This directly connects to the known spectral bias of coordinate MLPs toward low-frequency functions [1]. Consistently, in our matched control, replacing fixed Fourier features with a learned two-layer point MLP while keeping the same Transformer backbone decreases TU8 accuracy from 72.63% to 67.14%.

Therefore, the advantage is not due to a special algebraic property of persistent homology. Rather, Fourier features are particularly suitable for PDs because PDs represent information through point locations on the birth–death plane, while the relevant structure may occur across multiple spatial scales. Injectivity prevents point collapse, whereas the multi-frequency structure provides the more relevant inductive bias for learning such multi-scale coordinate structure.

[1] Tancik, M. et al. Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. NeurIPS, 2020.

Q5. Overall assessment.

Q5: "my assessment : I find the paper interesting, ..."

Thank you for the positive and constructive assessment. We will state the scope of the theoretical results more carefully.

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

The paper proposes Persistence Positional Encoding (PPE), which applies fixed random Fourier features to persistence-diagram coordinates and processes them using a standard Transformer.

The authors provide theoretical results on grid-based reconstruction and universal approximation, along with extensive experiments comparing PPE against existing persistence-diagram methods under a unified evaluation protocol.

Critical strengths

I think the problem addressed by the paper is interesting, particularly how persistence diagrams can be represented and processed by Transformer architectures without requiring specialized attention mechanisms.

I also appreciate the effort made in the experimental evaluation. The authors reproduce several persistence-diagram methods using the same splits, and validation-based model selection, which I think is useful considering the differences between experimental protocols in previous TDA papers.

Critical weaknesses

1- Limited methodological novelty and unclear contribution compared with previous works.

My main concern is that I could not identify what is fundamentally new in the proposed method. From what I understood, PPE mainly applies standard random Fourier features [1] to persistence-diagram coordinates and then uses a standard Transformer with learnable homology embeddings and CLS pooling.

Fourier-based multidimensional positional encodings for attention architectures were already studied by [2], while Persformer [3] already introduced Transformer-based learning directly on persistence diagrams and established universal approximation results.

The authors acknowledge that the Fourier feature mapping itself is not novel, but I still could not understand what the main methodological contribution is beyond combining these existing ideas and showing that this particular encoding works well experimentally.

Another point is that I could not understand what is actually gained from using a Transformer here. In Table 3, RFF with DeepSets [4] achieves 72.20% TU8 accuracy, compared with 72.63% for the Transformer, while RFF with sum pooling achieves 71.95%. These differences are relatively small, especially considering the quadratic computational complexity of self-attention. The experiments seem to demonstrate the importance of Fourier features instead of the importance of the proposed Transformer interface.

I also think the related work could better position the contribution within the broader literature on topological representations and positional encodings. For example, PiPE [5] already investigates the relationship between positional encodings and persistent homology, while TopNets [6] and TopoFormer [7] study other ways of integrating topological information into neural architectures. Although these approaches do not operate in exactly the same PD-only setting and are not necessarily direct baselines, I think their methodological and theoretical connections deserve a better discussion. PiPE is briefly mentioned in the appendix, but its relationship to the proposed method is not really examined.

Overall, the paper seems to propose a relatively simple combination of existing components, without clearly establishing why this should be considered a substantial new method rather than an empirical baseline.

2- Unclear theoretical contribution

I also have concerns about the theoretical contributions of the paper. Although the authors provide two theorems concerning grid-based reconstruction and universal approximation, I could not clearly understand what new insights these results provide for persistence-diagram learning or how they justify the proposed method.

In particular, I could not understand what these theoretical results establish specifically about the proposed Fourier encoding, considering that universal approximation for persistence-diagram Transformers was already studied in Persformer [3].

Overall, I think the authors need to better explain the novelty and significance of these theoretical results, both for the proposed method and for the broader topological ML community.


[1] M. Tancik et al. Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. NeurIPS, 2020.

[2] Y. Li, S. Si, G. Li, C.-J. Hsieh, and S. Bengio. Learnable Fourier Features for Multi-Dimensional Spatial Positional Encoding. NeurIPS, 2021.

[3] R. Reinauer, M. Caorsi, and N. Berkouk. Persformer: A Transformer Architecture for Topological Machine Learning. 2021.

[4] M. Zaheer et al. Deep Sets. NeurIPS, 2017.

[5] Y. Verma, A. H. Souza, and V. K. Garg. Positional Encoding meets Persistent Homology on Graphs. ICML, 2025.

[6] Y. Verma, A. H. Souza, and V. K. Garg. Topological Neural Networks go Persistent, Equivariant, and Continuous. ICML, 2024.

[7] M. J. Uddin, A. Tola, C. G. Akcora, and B. Coskunuzer. TopoFormer: Topology Meets Attention for Graph Learning. ICLR, 2026.

Questions for authors
  • What exactly is the main methodological novelty of PPE compared with combining previous Fourier-based positional encodings [1, 2] with persistence-diagram Transformers such as Persformer [3]? Could the authors also better explain how PPE differs from previous approaches to topological positional encodings, particularly PiPE [5] and TopoFormer [7]?

  • Given the results in Table 3, where RFF with DeepSets achieves 72.20% TU8 accuracy compared with 72.63% for the Transformer, what is actually gained from using self-attention?

  • What exactly do Theorems 1 and 2 contribute to the proposed method and to persistence-diagram learning more generally? In particular, what new theoretical insight does Theorem 2 provide compared with the universal approximation results already established for Persformer [3], and how does it justify the choice of Fourier features?

5 comments

We thank the reviewer for the careful reading. Several insightful questions are addressed in the General Response (GC1–GC3); below, we address the review-specific points.

W1-1 & Q1-1. Novelty and contribution.

W1-1: “My main concern is that …”

Q1-1: "What exactly is the main …"

Thank you for raising this concern. We agree that neither Fourier features nor Transformers are new in isolation, and we do not claim a fundamentally new architectural primitive. As detailed in GC1, our contribution is the controlled diagnosis and design principle that, among the choices tested, PD point representation is substantially more consequential than further Transformer specialization.

Relative to the works cited by the reviewer, Tancik et al. and Li et al. study Fourier encodings for spatial coordinates in broader neural or attention-based models, while PersFormer combines a learned pointwise embedding with additional PD-specific architectural choices. However, PersFormer did not fully resolve the PD-to-Transformer interface problem: PersFormer (i) shows slow and unstable training in our Fig. 2b, (ii) is reported by xPerT as “not trainable” on ORBIT5K, and (iii) remains below conventional PD vectorizations (PI/PL/Betti curve) on TU8. PPE instead freezes a random Fourier encoder and leaves the Transformer unchanged, allowing us to isolate representation from architectural specialization. Table 3 directly supports this distinction: Point MLP→RFF under the same Transformer improves TU8 from 67.14→72.63 (+5.49 pp), whereas RFF+DeepSets→RFF+Transformer changes only 72.20→72.63 (+0.43 pp). Thus, our contribution is not “RFF + Transformer” itself, but controlled evidence that an appropriate point representation can make a standard Transformer effective on PDs without PD-specific attention or pooling.

A second, important contribution is the unified evaluation itself. As emphasized by the ICML 2026 position paper [1], inconsistent inputs and evaluation protocols remain a major obstacle to reliable progress in topological machine learning. We therefore reevaluate Transformer-based methods, vectorizations, kernels, and learned PD representations under shared PD inputs, folds, search budgets, and validation-only model selection. Please see GC1 for further details.

[1] Castilla Rieso et al., ICML 2026.

W1-2 & Q2. What is gained from self-attention?

W1-2: "Another point is that …"

Q2: "Given the results in Table 3, …"

Thank you for raising this point. As discussed in W1-1 and GC1, our central question is how PDs can be made effectively usable by a standard Transformer, rather than whether Transformers are universally preferable to other set architectures. The Transformer is thus the fixed target of our problem rather than a design choice of our method.

W1-3 & Q1-2. Related work positioning.

W1-3: "I also think the related work …"

Q1-2: "Could the authors also better …"

Thank you for pointing this out. We will clarify these connections in the Related Work.

PiPE studies graph positional encodings and their relationship to persistent homology; importantly, its “PE” refers to graph PE, whereas PPE encodes the birth–death coordinates of PD points. TopNets integrates PH with GNNs/TNNs, while TopoFormer constructs topological tokens from graph filtrations. In contrast, PPE starts from an already-computed PD and studies how its birth–death points should be represented for a standard Transformer. The theoretical focus also differs: PiPE studies the expressivity relationship between graph PE and PH, whereas PPE analyzes direct point representation and fixed random Fourier encoding for PDs.

W2 & Q3. Unclear theoretical contribution

W2: "I also have concerns about …"

Q3: "What exactly do …"

Thank you for raising this point. As summarized in GC2, the two theorems support PPE’s specific representation choices rather than making broad claims about other PD methods.

Theorem 1 supports preserving individual PD points by showing that information lost through hard grid discretization cannot be recovered downstream. Theorem 2 differs from PersFormer’s universality result. PersFormer establishes universality for its learned PD Transformer, whereas our theorem shows that a randomly sampled Fourier bank can be fixed throughout training without losing universal approximation in our setting. This directly supports PPE’s use of fixed random Fourier features.

Together, the results provide representation-level justification for PPE’s two central choices, preserving PD points and using a fixed random Fourier encoding.

I thank the authors for their response and the additional experiments. I appreciate the clarification regarding PiPE, TopNets, and TopoFormer, and I consider this concern addressed.

However, my main concerns remain. The authors clarify that the main contribution is the empirical observation that point representation matters more than architectural specialization. While I find this observation interesting, I am still not convinced that applying an existing Fourier encoding to persistence diagrams provides a sufficiently significant methodological contribution.

Regarding the theory, I understand that Theorem 1 motivates point-level representations and that Theorem 2 establishes universality with fixed random frequencies. However, the first result does not explain the improvements over existing grid-free methods, and the second does not establish a particular advantage of Fourier features over other injective point representations.

Although I appreciate the experimental contribution, I believe the methodological and theoretical contributions remain limited. Therefore, I will maintain my original score of 2.

AuthorsOpenReview ↗

Response (1/2)

We thank the reviewer for the continued discussion. While we respect the reviewer’s assessment, we would like to further discuss what constitutes a meaningful methodological contribution, as we believe our work may be evaluated primarily through the novelty of its individual components rather than the findings it establishes.

On methodological novelty.

We agree that PPE does not introduce a new architectural primitive. However, we believe that the novelty of individual components and the methodological significance of a method are not necessarily the same.

A relevant precedent is the original Vision Transformer (ViT; Dosovitskiy et al., ICLR 2021). Its central idea was remarkably simple: tokenize images into patches and process them using an essentially standard Transformer. Interestingly, the original ViT even underperformed comparable CNNs when trained on ImageNet-scale data, requiring substantially larger pretraining datasets to become competitive. Under an evaluation criterion that places primary emphasis on the novelty of individual components and immediate performance gains over existing architectures, one could similarly regard the original ViT as having limited methodological significance. Nevertheless, its importance lies in demonstrating that a general-purpose Transformer can effectively process images without CNN-specific architectures, rather than in introducing fundamentally new architectural components or outperforming every existing method under all conditions.

We certainly do not claim that PPE has the same scope or impact as ViT. Rather, this example illustrates an important distinction: establishing a simple but effective interface between a new data modality and a general-purpose architecture can itself constitute a meaningful methodological contribution.

In our case, this interface connects a standard topological representation with a general-purpose architecture. Persistence diagrams are a standard representation of topological information, while Transformers are widely used across learning pipelines. An effective PD-to-Transformer interface can therefore make explicit topological information easier to incorporate into existing Transformer-based systems, rather than being tied to a specialized PD architecture.

Importantly, our work goes beyond demonstrating such an interface. We provide several complementary contributions:

  • Strong empirical performance: Despite its simplicity, PPE achieves the highest TU8 macro accuracy among all evaluated PD method–downstream configurations (73.22), outperforming the strongest prior PD–Transformer method by 3.90 pp. It also substantially improves upon prior PD–Transformer methods on ORBIT5K (+6.82 pp) and MCB-C (+9.46 pp).
  • A controlled methodological finding: Rather than merely combining existing components, we isolate and quantify the effects of point representation and downstream architecture. Across matched TU8 controls, changing the point encoding yields 5.49–6.47 pp, whereas changing the downstream module yields only 0.47–1.45 pp, showing that point representation is the dominant design variable among those tested.
  • Theoretical justification: We establish an irreducible reconstruction limitation for uniform grid averaging and prove universal approximation for a Transformer with fixed random Fourier frequencies. These results justify specific design choices rather than claiming superiority over all alternative representations.
  • A unified evaluation protocol: We systematically reproduce and reevaluate methods across PD-learning families using identical PD inputs, shared splits, and validation-only model selection. This directly addresses the lack of experimental comparability identified as a major obstacle in recent topological ML literature.

We additionally introduce a sub-cell discrimination benchmark, where PPE outperforms the strongest competing PD–Transformer method by 17.86 pp, providing further evidence of the practical importance of preserving point-level information.

Comments on the paper

2 comments

We thank all reviewers. Before the individual responses, we briefly recap three key points (GC1–GC3) to provide context for our responses.

(1) GC1. Positioning and Contribution

Positioning. Our question is how Transformers can effectively use persistence diagrams (PDs) as explicit topological information. This remains unresolved: prior Transformer-based methods (xPerT/PersFormer/MST) achieve 69.32/66.06/64.68 on TU8, below conventional PD vectorizations (PI/PL/Betti curve) at 71.67/71.17/71.01 (Table 2). The xPerT paper reports PersFormer as “not trainable” on ORBIT5K, consistent with the slow, unstable training shown in our Fig. 2b. FILTR [1] (CVPR 2026 Oral) reports that Point-BERT CLS reaches only 57.2% for connected components and 25.9% for genus on DONUT, suggesting that pretrained 3D representations may not encode global topology well. These results motivate better interfaces between explicit topology and general-purpose neural backbones.

Why this matters. Transformers are general-purpose backbones across modalities. A simple PD-to-Transformer interface would let existing pipelines use explicit topology without domain-specific architectures. Our goal is not to build a new SOTA model, but to make explicit topological information effectively usable by standard Transformers.

Novelty. Transformers and Fourier features are established, and our novelty lies in the problem diagnosis and design principle. Prior PD–Transformer work develops PD-specific pipelines through specialized tokenization, embedding, attention, or readout. PersFormer modifies point embedding, encoder connectivity, and pooling, xPerT introduces pixelization, patch tokenization, and sparse token selection, and MST uses multiset-specific attention and aggregation. We take a different direction. We leave the Transformer unchanged and focus only on how each PD point is encoded. This isolates representation from architecture. Table 3 shows that representation is the more consequential choice. Point MLP→RFF under the same Transformer improves TU8 from 67.14→72.63 (+5.49 pp), and raw coordinates→RFF under DeepSets improves 64.92→72.20 (+7.28 pp), whereas with RFF fixed, DeepSets→Transformer changes only 72.20→72.63 (+0.43 pp). Thus, in the tested settings, point representation matters substantially more than downstream architectural specialization. We do not claim that self-attention is necessary. Rather, an appropriate representation lets a standard Transformer use PDs effectively without PD-specific attention or pooling.

Contribution. We contribute (i) controlled evidence that point representation is the dominant design variable tested, (ii) a minimal PD-to-Transformer interface without PD-specific Transformer modifications, and (iii) a unified evaluation protocol with the same PD inputs, folds, search budgets, and validation-only selection, addressing the lack of comparable experimental standards highlighted in [2] (ICML 2026 position). PPE provides a simple Transformer-native baseline for future PD–Transformer methods.

[1] Martinez, L. and Ovsjanikov, M. FILTR: Extracting Topological Features from Pretrained 3D Models. CVPR, 2026.

[2] Castilla Rieso et al., “Position: Topological Machine Learning Cannot Progress without Experimental Standards,” ICML 2026.

(2) GC2. Theorems 1–2

Each theorem supports our specific design choices of direct point representation and fixed Random Fourier Feature encoding, rather than serving as a general criticism of other approaches.

  • Theorem 1 → preserve points rather than hard-binning. Uniform cell averaging leaves a worst-case reconstruction error at every resolution and for every decoder.
  • Theorem 2 → fixed random frequencies preserve expressivity. PPE samples frequencies once and keeps them fixed. With at least three frequencies, almost every draw retains universal approximation.

(3) GC3. Experiment for Stability

Experiment. ORBIT5K, 1,500 test clouds (3 folds)

  • Raw: uniform noise in [−ε,ε][-ε,ε] added to the cloud.
  • PD: (b,d)(b,d) perturbed directly.
  • Entries: accuracy/predictions changed(%), where predictions changed(%)=(changed predicted classes/all samples)×100.
  • Bold marks changed ≥ 10%.
NoiseεPPExPerTMSTPI
Clean–91.20 / 081.67 / 082.13 / 080.00 / 0
Raw1e-490.67 / 2.1381.67 / 2.9381.73 / 1.6780.07 / 0.07
Raw1e-391.00 / 3.6082.47 / 5.4081.33 / 3.0079.80 / 2.13
Raw1e-284.67 / 10.2078.67 / 15.7377.33 / 15.2776.93 / 13.07
PD1e-491.17 / 0.1781.70 / 1.8781.80 / 1.4379.57 / 10.27
PD1e-390.97 / 1.9081.67 / 9.2778.10 / 10.9746.73 / 47.80
PD1e-235.10 / 66.0350.43 / 48.7020.23 / 79.8723.20 / 78.27

Overall, PPE is empirically stable to small perturbations, showing competitive robustness even against PI, a standard PD vectorization with an established stability guarantee.

AuthorsOpenReview ↗

Summary of Revisions and Responses

Dear Reviewers, Area Chairs, Senior Area Chairs, and Program Chairs,

As the author–reviewer discussion period approaches its end on November 18 at 11:59 pm AoE, we would like to express our sincere gratitude for the time and effort you have dedicated to reviewing our paper. We have carefully considered all comments and uploaded a revised manuscript, with the changes highlighted in orange. The revision incorporates the clarifications, analyses, and additional results provided in our responses, and we summarize the main changes below (line numbers refer to the revised PDF).

Claims and scope

  • Benchmark claim: Refined the benchmark statement in the Abstract and clarified the role of non-Transformer baselines as reference points (Abstract, lines 026–027; Sec. 4.2, lines 348–350).
  • Scope of Theorems 1–2: Clarified the role of the two theorems in motivating the design choices of PPE and their intended theoretical scope (Sec. 3, line 201; Sec. 3.2, lines 233–234).
  • Role of the Transformer: Clarified the role of the Transformer within PPE and its relationship to the proposed point representation (Sec. 4.3, lines 424–425).

Stability and sensitivity (Appendix J)

  • Stability: Added Proposition 3, which extends the pointwise bound to the complete PPE output and, under an explicitly stated assumption on the trained model, to raw-input perturbations, together with the design rationale for retaining near-diagonal points (App. J, lines 1477–1517; Sec. 5, lines 484–486).
  • Empirical sensitivity analysis: Added the raw-input and PD-coordinate perturbation results (Table A13, lines 1519–1535).
  • Frequency scale: Added the σ sweep analysis to characterize the effect of frequency scale (Table A14, lines 1538–1551; Sec. 4.3, lines 431–432).
  • Why Fourier features: Added a frequency-content explanation that further clarifies the role of Fourier features in the representation (App. J, lines 1554–1566).

Positioning and related work

  • Motivation: Strengthened the motivation by noting the complementary graph-distinguishing capability of persistent homology (Sec. 1, lines 039–040).
  • Related work: Clarified the relationship to PiPE, added TopNets and TopoFormer (App. B, lines 944–950), and further positioned our contribution relative to established pointwise embeddings (App. B, lines 885–887).

The revised main text remains within the 10-page limit for the discussion phase.

We understand that some reviewers may not have had the opportunity to reply to our latest responses. We would be grateful if you could take our responses and the corresponding revisions into account in your final assessment. If any concerns or questions remain, we would be happy to address them before the discussion period ends.

Thank you again for your time and valuable feedback.

Best regards,
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.