- 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
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.
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.
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 and , 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 . A bottleneck matching has two cases: (i) off-diagonal points matched to off-diagonal points, where every feature with persistence must be matched off-diagonal; and (ii) points matched to the diagonal, which necessarily have persistence .
(i) From pointwise to complete matched-diagram stability (persistent off-diagonal features).
Let contain matched points with label-preserving bijection satisfying
For
for RFF frequencies , satisfies
Since
with input projection and label embedding , which cancels for matched labels; for token matrix (Frobenius norm) and induced permutation ,
PPE has no sequence positional encoding, so does not affect the CLS output. The fixed Transformer/readout is Lipschitz on this bounded fixed-cardinality domain, with finite -point constant ,
where is the PPE logit vector.
(ii) Low-persistence features matched to the diagonal.
If , every point matched to the diagonal satisfies
hence has persistence at most . 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 off-diagonal points, suppose uniformly, for some , that the trained model satisfies
Define
Under this condition,
Indeed, off-diagonal matched pairs are controlled by (i), while each diagonal-matched point contributes at most . Since each diagram has at most off-diagonal points, their total contribution is at most .
Combined with standard PH stability for the chosen filtration,
where is the raw metric and 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 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.
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.
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.
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.
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.
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.
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).
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.
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.
Thank you for the suggestion. Due to space constraints, please refer to GC3 for our response to this point.