acceptodds
Under review as a conference paper at ICLR 2027

COMPLEX: A Closed-Form Certified Embedding of Multiparameter Persistence Modules

Abstract

Every multiparameter persistence vectorization we know of carries a one-sided Lipschitz upper bound and nothing below it: without a lower gauge there is no sense in which the features are faithful, and no per-prediction guarantee can be built on them. This paper supplies the missing side. COMPLEX is a closed-form, training-free embedding of multiparameter modules—slice the module along a fixed near-diagonal net, embed each slice barcode by the certified PLACE/PALACE landmark map, concatenate. Under a checkable witnessing-slice coherence condition, holding on 100% of audited pairs on Orbit5k, a single slice carries a closed-form lower gauge: separated modules stay separated in the embedding. With the standard upper bound this gives, to our knowledge, the first two-sided distortion bound for a multiparameter feature map, making faithfulness measurable. Measuring it, we find the floor tight within a small factor of realized distances yet operationally local: an RBF-SVM reaches 91% where 1-NN reaches 78% on the same features. Local per-prediction certification therefore fails for a structural reason common to every landmark embedding whose lower gauge is witnessed by one coordinate. With no learned embedding and no held-out calibration—only a cross-validated SVM head—COMPLEX matches the best published accuracy on Orbit5k (91.95%) and exceeds it on Orbit100k (92.98%), level with Euler-characteristic surfaces and above transformers and graphcode. On graphs it exceeds GRIL on all four shared molecular benchmarks with one fixed configuration, including the only multiparameter method to clear COX2's majority baseline by more than three points. Closed-form selection—of the landmark radius, the kernel (certificate-preserving), and the bifiltration set—buys further accuracy; gradient-shaped adaptation buys no reliable gain.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
multiparameter persistence, topological data analysis, certified embeddings, distortion bounds, graph classification, representation learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_1XuqOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a training-free vectorization for multiparameter persistence modules based on applying PLACE/PALACE embeddings across a fixed set of 1D slices and concatenating the resulting features. The authors derive upper and lower bounds under a coherence assumptions sinilar to one in the 1D case, and evaluate the method on point-cloud benchmarks and molecular graph datasets, where it achieves competitive or improved classification accuracy relative to several existing topological baselines.

Critical strengths
  • Competitive empirical performance with a simple, training-free representation. COMPLEX matches several topological baselines across the Orbit and molecular graph benchmarks, despite using a fixed representation followed by a standard SVM.

  • A useful theoretical characterization of when the embedding is faithful. Beyond the standard Lipschitz upper bound, the paper provides a conditional lower distortion guarantee through the witnessing-slice coherence condition. Although this condition is closely related to the existing 1D PALACE assumption, making explicit when the sliced multiparameter construction preserves separation is still useful for understanding the behavior and limitations of the method.

Critical weaknesses
  • Limited theoretical novelty of the lower bound. The main lower-distortion result appears to be obtained by selecting a slice that witnesses the multiparameter separation and then assuming the existing 1D PALACE coherence condition on that particular slice. Thus, the result seems closer to a conditional lifting of the known 1D guarantee than to a genuinely new multiparameter coherence or lower-bound mechanism. In my view, this substantially narrows the claimed theoretical novelty. Unfortunately, I do not see a substantial conceptual leap beyond the existing 1D PALACE argument, despite the paper's claim to provide the first lower-distortion guarantee for multiparameter persistence.

  • Overall novelty appears limited. Beyond the lower-bound argument discussed above, I am not convinced that the paper as a whole introduces sufficient methodological or conceptual novelty.

  • Gap between the theoretical guarantee and the empirical validation. The experimental configurations do not generally enforce the assumptions required for the lower-bound guarantee, and the paper instead relies on post-hoc coherence audits. Moreover, the classification improvements over the strongest existing baselines are relatively modest. Therefore, while the empirical results are competitive, they do not clearly demonstrate that the proposed theoretical faithfulness mechanism is responsible for the observed performance.

Questions for authors
  1. The lower-distortion guarantee appears to follow by choosing a slice that witnesses the multiparameter distance and then imposing the existing 1D PALACE coherence condition on that slice. Could the authors clarify precisely what part of this argument is genuinely multiparameter and conceptually new, rather than a conditional lifting of the 1D PALACE result? This point would materially affect my assessment of the theoretical novelty.

  2. The empirical improvements over the strongest baselines are generally modest, while the assumptions required for the lower-distortion guarantee are not enforced during the experiments. Could the authors provide stronger evidence that the reported gains are statistically meaningful and/or demonstrate that satisfaction of the witnessing-slice coherence condition is actually related to improved downstream performance?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 1Xuq

Q1 (what is multiparameter and new). You are right that the per-slice step is PALACE's and we reuse it deliberately as the certified primitive. The multiparameter content is in four places, none of which exists in one parameter:

  1. The witnessing-slice mechanism with scale cancellation (Thm 2.3). The floor is not "impose PALACE's coherence on a chosen slice": the slice weight ω_s is a realized scale that enters the active-set condition (r ≥ τ/(4ω_s)) and cancels in the bound, giving a floor that is independent of the slice angle. This is why a fixed near-diagonal net suffices and why the gauge is for the sliced matching distance d_matS rather than for any single barcode distance. Reviewer Xvs5 singles out exactly this cancellation as the contribution.
  2. Coherence is a property of a pair across a family of slices, not of a fixed one-parameter module: the hypothesis is that some slice witnesses the separation at its own scale, and the theorem does not need to know which (Definition 2.4). In one parameter there is one slice and nothing to witness.
  3. The bounded-complexity restriction is forced, not chosen. Already the space of one-parameter diagrams of unbounded size has no coarse embedding into a Hilbert space (Bubenik–Wagner 2020; Zava 2025, the latter cited in the submission), and the sliced module space inherits this slice by slice, so no lower gauge of the form of Theorem 2.3 can exist without a complexity bound. The guarantee is therefore stated on R_N (Remark B.3), and the bar budget N is the explicit price. We will add a sentence making this reasoning explicit in Section 2.
  4. Near-diagonal net optimality: the per-slice gauge is maximized at the diagonal and degrades as the slice tilts (eq. (9) makes the required radius τ/(4ω_s) unbounded as ω_s → 0), which determines the net; the one-parameter theory has no net.

We will state (1)–(4) as the contribution list and say that the per-slice embedding is PALACE's.

Q2 (significance; coherence and performance). Significance (identical folds, 3 seeds × 10 folds, Nadeau–Bengio corrected paired test): single-bifiltration COMPLEX (91.74) vs. ECS+XGB re-run on the same bifiltration +2.50 (p = 2.7e-6), vs. HT +4.52 (p = 1e-9), vs. one-parameter PALACE +4.38 (p = 1e-10); the two- vs. one-parameter gain at α = 1.75 is +3.89 (30/30 folds, p = 7.7e-9). Against published numbers (Welch from reported mean ± s.d.) COMPLEX ties ECS (−0.06, p = 0.90), PALACE (+0.44) and Persformer (+0.54, p = 0.30) and beats Graphcode (+3.24, p < 0.01); our in-protocol ECS re-run (89.2 over 3 seeds) is 2.6 points below the published 91.8, so the revision says "ties the best published results", not "beats ECS". On graphs the multi-bifiltration configuration beats the single bifiltration on DHFR (+3.1, p = 0.03), COX2 (+3.9, p = 0.03) and NCI1 (+1.6, p = 0.006), not on MUTAG or PTC; against cited GRIL the gaps (+1.2 to +2.6) are not significant, and the revision says so.

Coherence and accuracy: we tested, for every test point, the 20 nearest cross-class training modules at the pair's own scale. The raw association runs the wrong way (incoherent pairs are more accurately classified) and is a scale confound: incoherent nearest pairs have median d_matS 1.45 vs. 0.32 for coherent ones. Coherence holds on 100% of pairs in the bottom half of cross-class separation, where the floor is the binding quantity, and fails only in the top deciles, where accuracy is 98–99% anyway. Stratified by separation decile the association vanishes (Cochran–Mantel–Haenszel p = 0.91). What does predict accuracy is the theory's own quantity: SVM accuracy rises monotonically with the nearest cross-class d_matS, from 79% in the lowest decile to 99% in the highest, at both α. So the mechanism behind the lower bound (small separation is where a floor is needed, and there it holds) is visible in the data, but we do not claim that coherence causes the accuracy gain, and the revision says the gain over the best published numbers is a tie.

Thank you for the detailed responses and additional experiments. I appreciate the author efforts to address the concerns. The new experiments provide substantially stronger evidence for the empirical value of the multiparameter construction. Furthermore, I appreciate the authors' decision to acknowledge these limitations and revise their claims accordingly.

I also acknowledge that the witnessing-slice argument and scale cancellation introduce nontrivial technical considerations specific to the multiparameter setting. However, my main concern regarding theoretical novelty remains. Although the resulting lower-distortion statement is formulated for multiparameter persistence, its central mechanism still appears to rely on applying the existing 1D PALACE guarantee to a witnessing slice under an analogous coherence condition. I view the scale cancellation and slice selection as useful extensions, but I am not yet convinced that they constitute a sufficiently substantial conceptual advance.

Overall, the rebuttal has improved my assessment of the paper's empirical rigor and clarified its contributions. Nevertheless, I remain concerned about the degree of methodological and theoretical novelty relative to existing work. I therefore maintain my current recommendation of weak rejection.

AuthorsOpenReview ↗

On what Theorem 2.3 does that a one-parameter guarantee cannot

Thank you for reading the responses and for stating what moved and what did not. On the one point that remains we disagree, and we want the record to show why.

"Applying the existing 1D PALACE guarantee to a witnessing slice" is not what Theorem 2.3 does. The one-parameter guarantee certifies a bottleneck distance at a fixed scale on one barcode. Theorem 2.3 certifies the sliced matching distance, a maximum over slices of barcode distances each weighted by its slice's realized scale, for a pair of modules; it does so without knowing which slice attains the maximum, under a condition (Definition 2.4) that is existential over the family of slices at each slice's own scale; and the floor it yields, τ w_min/4, is the same for every slice. Applying the one-parameter bound at a fixed scale on each slice does not give this: it gives a floor proportional to the weight of the witnessing slice, which degrades to zero as the slice tilts. The theorem's floor does not degrade because the active-set condition is posed at the slice's realized scale τ/(4ω_s) and the weight cancels against it in the bound. That cancellation is the mechanism, not an extension of one. It is why a fixed near-diagonal net suffices, why the lower gauge is for d_matS rather than for any single barcode distance, and why no one-parameter result contains the statement.

We say in the revised contribution list, in the same breath, that the per-slice embedding is PALACE's; we reuse it deliberately as the certified primitive, and the multiparameter content is the four items of our earlier reply. The committee can weigh the same facts we have put here. We are grateful for the review; the paper is more honest for it.

Thank you for the clarification. I acknowledge that the witnessing-slice argument and scale cancellation yield a valid multiparameter extension of the existing 1D PALACE guarantee. However, the underlying mechanism still appears to follow from applying the known componentwise bound at a rescaled threshold and lifting it through the weighted concatenation. While this is a useful technical extension, I remain unconvinced that it represents a sufficiently substantial conceptual leap beyond the existing framework.

Nevertheless, I appreciate the authors' considerable efforts to address the concerns, conduct additional experiments, correct earlier results, and improve the paper's rigor and presentation. I believe the work has merit and could make a valuable contribution in a more specialized venue focused on topological data analysis. However, my overall assessment remains unchanged.

Reviewer_Xvs5OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

COMPLEX concatenates weighted PLACE/PALACE landmark embeddings of barcodes obtained on a finite net of positive-slope slices. The upper stability bound follows by stacking per-slice bounds. Its lower separation gauge is conditional on a witnessing slice having a landmark coordinate with sufficient change; a separate, strong within-module condition implies this coherence. The paper connects the representation to deterministic nearest-neighbor stability and population-nearest-centroid agreement, and reports competitive SVM accuracy on Orbit and TU datasets. Extensive ablations expose the low firing rate of local certificates, weaker nearest-centroid accuracy, and limited gains from learned filtrations. The empirical pipeline uses truncated barcodes and several variants of the formal embedding; the paper explicitly acknowledges that its probabilistic certification guarantees are not established for the same-sample landmark selection used in the experiments.

Critical strengths
  1. A precise conditional construction. Theorem 2.3's cancellation of the slice weight against the realized scale gives an explicit angle-independent floor, while Proposition B.7 supplies a separate sufficient condition for coherence (pp. 4–5, 15–17). The paper also clearly distinguishes the finite-net distance from the continuous matching distance, and pairwise separation from correctness or SVM margin.

  2. Useful empirical breadth and negative results. The reported SVM results are competitive across point-cloud and molecular-graph benchmarks, with radius, degree-pooling, slice, kernel, and filtration ablations. The weak nearest-centroid/1-NN performance, failure of reliable gradient gains, and OGB deficit are reported rather than hidden (pp. 7–9, 23–34). This provides a potentially useful practical study of the representation.

Critical weaknesses
  1. The advertised certification benefit is not demonstrated for the evaluated pipeline. Definition 2.4 assumes a witnessing coordinate separation, from which Theorem 2.3's norm floor follows by weighting and stacking; Proposition B.7 supplies a stronger geometric sufficient condition that the configurations do not satisfy (pp. 4–5, 15–17). Direct coherence could still suffice, so failed admissibility is not itself a counterexample. The empirical audits sample pairs rather than establish every pair required by a certificate. The nearest-centroid probability guarantees further assume an independently fixed configuration, whereas landmarks are selected from the same sample (p. 6), and Table 8 substitutes estimates of population variance/direction quantities without controlling that estimation error (pp. 21–22, 30–31). The paper acknowledges these gaps, but threshold-firing rates are consequently not validated probability guarantees. Also, a small radius from the chosen triangle-inequality SVM bound cannot rule out every tighter global or local Lipschitz certificate (pp. 8–9, 28–29).

  2. The empirical evidence is difficult to verify or attribute across protocols. The supplied ZIP has only manuscript source, bibliography/styles and figures; it does not include the named experiment/audit scripts. Important calibration, graph construction, tuning and training choices cannot be recovered sufficiently to reproduce the results (p. 7; Appendix E). Many representation variants and ablations are disclosed, but their precise association with each table and theorem remains hard to audit.

Questions for authors
  1. A proof-valid certification experiment: Can Table 8 be instantiated with a configuration fitted independently of the centroid-estimation sample and valid confidence control for the unknown variance/direction quantities? Please separate any such guarantees from heuristic threshold firing and state the resulting coverage and confidence.

  2. Exact hypotheses and reproducibility: Can you provide the anonymous experiment/audit scripts, configurations and fold/seed records, together with a table mapping each experiment to weighted/unweighted, capped/uncapped, pooled/per-degree and truncated variants? For accepted 1-NN predictions, does the audit verify all relevant differently labeled pairs, rather than only sampled coherence? Equation (13) uses a cross-class minimum whereas p. 31 describes a quantile. Which condition is actually certified? Direct verified coherence may replace the stronger sufficient admissibility condition; please identify the route used.

  3. Scope of the negative result: Can the claim that no Lipschitz route certifies the accurate head be restricted to the tested upper bounds, unless tighter global/local bounds are excluded? The dual-coefficient triangle bound is not the exact Lipschitz constant, so its failure alone is insufficient for the universal conclusion.

5 comments
AuthorsOpenReview ↗

Response to Reviewer Xvs5

Thank you for the most precise review this line of work has received; your reading of Theorem 2.3 (the slice weight cancels the realized scale) is the point of the paper.

W1 / Q1 (proof-valid certification). We implemented the experiment you asked for: landmarks, radii, τ and a pilot centroid are fitted on one stratified half of each training fold; centroids on the other half; the variance term is bounded by an empirical-Bernstein bound with an almost-sure range from the landmark geometry (no plug-in); the radius by Pinelis; the direction cross-term by the triangle inequality; total confidence δ = 0.05 with an explicit split; calibration modules kept out of both halves and the test set. Result: 0% of test points certified on all 11 benchmarks, at both split fractions (0.5, 0.2). On the same splits the plug-in version of Table 8 still "certifies" 31% of NCI1 (fit fraction 0.5) and 52% of NCI109 (fit fraction 0.2). The valid certificate fires on a synthetic population in our unit tests, so the 0% is not a bug: the worst-case range term (N bars on one landmark) enters at order 1/n and dominates the variance at these class sizes. To test whether this is a sample-size limit we repeated the valid certificate on Orbit100k's diagrams at class sizes m ∈ {1,000, 2,000, 5,000, 10,000, 20,000} (3 folds, seed 0, same split rule): valid coverage is 0% up to m = 5,000, 0.6% at 10,000 and 12.1% at 20,000 (84.7% accuracy on the certified points), while the median ratio of the certified gap to the required gap rises 0.002, 0.008, 0.042, 0.116, 0.353. The certificate becomes operational at about 20,000 modules per class; Orbit5k's 900 per class is 20× short. The revision reports this curve in place of Table 8's plug-in rates. We agree this means the Table 8 rates are not validated probability guarantees, and the revision says so. We also agree that Prop. B.7's failure is not a counterexample to coherence: the revision now reports direct verification (Q2d below) and drops the admissibility route from the experiments.

Q2b (every pair or a sample). The separation hypothesis of Corollary 3.2 was already exhaustive (sep[i] is the exact minimum over all differently-labelled training modules). Coherence was not; it relied on the sampled audit. We added an exhaustive check: for every accepted test point, Definition 2.4 is tested on every cross-class pair (M_{i*}, M_j). Over 204,020 pairs, 99.32% are coherent under the exact definition (98.68% in the realized-distance form). After a scale correction we found while re-checking (the released PALACE code scales embeddings by 2^{-3/2}, a constant absent from Definition 2.1, and our audits compared code-scale distances with the paper-scale floor; the tightness numbers become 12.5× min / 29.8× median) the corollary accepts 0–0.5% of test points at K ≥ 16; the K = 8 cells reaching 2.5% (p. 31), which the submission already discounts as noise, are near-duplicate modules across classes and are dropped from the revision. Of the verified acceptances, about half are wrong predictions, which is consistent with the corollary (it certifies a 1-NN margin, not correctness) and is stated.

Q2c (minimum vs. quantile). Both are certified; the prose on p. 31 was imprecise. In the sweep, sep[i] is the exact minimum of d_matS from training module i to every differently-labelled training module (no sampling), τ runs over the quantiles {0, .25, .5, .75, .9} of these minima, and a test point is accepted only if its nearest neighbour's sep is ≥ τ. So at every row the separation hypothesis of Corollary 3.2 holds exactly for each accepted point, and Eq. (13), the global minimum, is the 0-quantile row at which every training point is eligible. The revision states the acceptance rule in these terms and reports the rows as "τ at the q-quantile of the exact cross-class minima".

Q2d (which route). Direct verified coherence, exhaustively (above). Admissibility (Prop. B.7) is retained only as a sufficient condition in the appendix and is not used by any experiment.

Q3 (scope of the negative result). Agreed and rescoped: "no Lipschitz route certifies the accurate head" (Sections 3 and 5) now reads "the dual-coefficient Lipschitz bound L_ab = e^{-1/2} σ^{-1} Σ_i |α_i^{ab}| does not certify it, and no constant down to L = 1 would at these margins"; we state that L_ab is an upper bound on the true Lipschitz constant, not its value, so a tighter global or local Lipschitz certificate is not ruled out by our measurement.

W2 / Q2a. The supplementary ZIP omitted the scripts by mistake. An anonymized repository with every experiment and audit script, cluster job files, seeds, per-fold accuracy files, and the table mapping each table/theorem to its variant (weighted/unweighted, capped/uncapped, pooled/per-degree, bar budget N) will be posted in a follow-up comment within the discussion period.

Thank you for the response, which addressed some of the concerns.

I would appreciate a clarification about Corollary 3.2. The corollary assumes that the configuration is τ\tau-admissible, while the paper states that the experimental configurations do not satisfy this assumption. Your response explains that the experiments instead verify coherence directly for the relevant cross-class pairs.

Could you clarify whether Corollary 3.2 can be stated under these directly verified conditions, together with the required active-set assumptions? If so, please also clarify whether the reported 1-NN coverage counts only predictions for which all the conditions needed for the guarantee have been verified.

AuthorsOpenReview ↗

Corollary 3.2 under the verified conditions, and the verified coverage

Thank you; this is the right question, and the answer is yes on the first part and "with one correction to the table" on the second.

Can Corollary 3.2 be stated under the directly verified conditions? Yes, and it should be: its proof (Appendix C) uses only Theorem 2.3 pairwise and Proposition 2.2, never admissibility. Admissibility is the hypothesis of Proposition B.7 only, the sufficient condition that derives coherence; where it appears elsewhere (Remark B.5, and the certifiability filter of Section 5, which builds admissible nets) it is a design target, not a hypothesis of a stated guarantee. The word "τ-admissible" in the statements of Corollary 3.2 and of Proposition D.1 is a leftover of the derivation route. The corollary holds verbatim with the hypotheses that are actually used, each of which is checkable: (a) separation: d_matS(M_{i*}, M_j) ≥ τ for every j with y_j ≠ y_{i*}; (b) coherence of each such pair in the sense of Definition 2.4, which includes the active-set requirement: the witnessing landmark k* must lie in A_{s*}(τ), i.e. r_{k*} ≥ τ/(4ω_{s*}) on the witnessing slice s*; (c) the floor: with equal landmark weights w = K^{-1/2}, Theorem 2.3's bound for a coherent pair is τ w_{k*}/4 = τ/(4√K), which needs only the witnessing slice's active set to be nonempty, and (b) implies that; the "all active sets nonempty" clause of Theorem 2.3 is needed only to define ρ_ν as a minimum over slices, and with equal weights it plays no role; (d) ε < ρ_ν(τ)/2. The PDF uploaded today states Corollary 3.2 and Proposition D.1 with (a)–(d) in place of admissibility, and Remark B.5 says which hypothesis is enforced by construction (a, by the acceptance rule: sep_{i*} is the exact minimum over all other-class training modules), which is computed (the active sets), and which is verified by audit (b, exhaustively).

Does the reported coverage count only fully verified predictions? Not as computed in the sweep, and we correct it. The acceptance rule in cert_1nn.py enforces (a), (c) and (d); (b) is then checked exhaustively for every accepted point (every cross-class pair of its neighbour, under the exact Definition 2.4 including the active set) and reported in the logs as "verified cover". Across all runs in the release (cert1nn_*_v3.txt): 212,865 pairs checked, 99.33% coherent. At K ≥ 32, every accepted point passed (all 26 accepted points, 26 of 26 neighbours fully coherent), so the reported 0–0.5% there is the verified coverage. At K = 16, 22 of the 31 accepted points passed (18 of 23 distinct neighbours); the five neighbours that failed carry the other nine accepted points, in five cells, whose verified coverage is 0 (they had read 0.25–0.75%), while the 0.5% cells are fully verified, so the verified range at K ≥ 16 is 0–0.5%. At K = 8, which the revision had already dropped as degenerate, only 12 of 80 accepted neighbours pass, which is a further reason to drop it. The uploaded PDF reports the verified coverage in Section 4 and in the settings table, with the K = 16 cells corrected, and the selective accuracy on the verified sets (one or two points per cell; 1.0 in some cells, 0.0 in others), which is consistent with the corollary certifying a 1-NN margin rather than correctness.

Reviewer_5LuNOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper proposes COMPLEX, a training-free embedding of multiparameter persistence modules that combines landmark features from a fixed collection of one-parameter slices. It provides stability and conditional separation guarantees, and evaluates the resulting features on point-cloud and graph classification tasks.

Critical strengths

The construction is simple and computationally efficient. The reported classification results are promising, and the comparison with the single-parameter approach helps illustrate the benefit of using multiple slices.

I appreciate the discussion of the relationship between classification accuracy and certification. The paper reports limitations of the certificates alongside the empirical results, which provides useful context for understanding the method.

Critical weaknesses

The scope and novelty of the theoretical contribution could be explained more clearly

The practical implications of certification remain somewhat unclear. The strongest classification results and the certification results involve different settings, making their relationship difficult to assess. The experimental comparison would also benefit from broader coverage of related methods and clearer discussion of differences in evaluation protocols

Questions for authors

Could the authors give a more intuitive explanation of the coherence assumption and clarify how the guarantees compare with prior work?

Could the authors clarify which experimental settings satisfy the theoretical assumptions and what the certificates provide in those settings?

Could the authors expand the discussion of related graph baselines and qualify the comparative claims where evaluation protocols differ?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 5LuN

Q1 (intuition for coherence; comparison to prior guarantees). The embedding sums hat functions over bars, so two different barcodes can produce the same coordinate (sum-pooling is many-to-one). Coherence says: for a separated pair of modules, there is at least one slice and one landmark on it whose coordinate actually sees the separation, at that slice's own scale. Every prior multiparameter vectorization (landscapes, images, signed measures, GRIL, ECS) has an upper Lipschitz bound only: nearby modules map nearby, but distinct modules may collapse. Theorem 2.3 is the first lower bound, and it is conditional because without coherence collapse can happen (Remark B.6 gives the mechanism). The revision adds this paragraph to Section 2.

Q2 (which settings satisfy the assumptions, and what is certified). The revision adds a table with one row per experiment: configuration, whether coherence was audited (sampled or exhaustive) and at what rate, whether Corollary 3.2 or Theorem 3.1 applies, and what is certified. In brief: on Orbit5k at α = 4 with the radius cap, coherence holds on 100% of audited pairs and 99.3% of 204,020 exhaustively checked pairs, and the deterministic two-sided bound (Prop. 2.2 + Thm 2.3) holds on every coherent pair with the realized distance at least 12.5× the floor; the per-prediction 1-NN certificate (Corollary 3.2) accepts under 0.5% of test points; the probabilistic nearest-centroid certificate, in the proof-valid form a reviewer asked for, certifies 0% on every benchmark (details in our reply to Reviewer Xvs5), and we now label the submitted Table 8 as a heuristic diagnostic. The accurate head (RBF-SVM, 91–92%) is not certified; the paper's claim is faithfulness of the features (a measurable two-sided bound), not certified accuracy.

Q3 (graph baselines and protocol differences). The revised Table 2 adds a protocol column (cited vs. re-run under our folds; node-label use; number of folds and seeds) and qualifies each comparison accordingly: against cited GRIL our gains (+1.2 to +2.6) are not significant; against our own single-bifiltration run the multi-bifiltration configuration is significantly better on DHFR, COX2 and NCI1 only. COX2 is degenerate (every multiparameter method, including D-GRIL, reports the majority rate 78.16); we keep the observation that COMPLEX clears it by more than three points but no longer call it a win over the field.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Global response: re-runs, corrections, and what the paper now claims

We thank the reviewers. Three of their requests led us to re-run or re-check our own experiments; we report the outcome plainly, including where it weakens the paper.

1. Code and variant map (Xvs5 W2, Q2a). An anonymized repository with every experiment and audit script, cluster job files, seeds, per-fold accuracy files, and a table mapping each table/theorem to its variant (weighted/unweighted, capped/uncapped, pooled/per-degree, bar budget N) will be posted in a separate comment within the discussion period; the supplementary ZIP omitted the scripts by mistake, and the revised PDF's Appendix E reproduces the mapping table.

2. A proof-valid version of Table 8 certifies nothing (Xvs5 Q1). We implemented the experiment the reviewer asked for: landmarks, radii, τ and a pilot centroid are fitted on one stratified half of each training fold; centroids on the other half; the variance term is bounded by an empirical-Bernstein bound with an almost-sure range from the landmark geometry (no plug-in); the radius by Pinelis; the direction cross-term by the triangle inequality; total confidence δ = 0.05 with an explicit split. Calibration modules are kept out of both halves and the test set. Result: 0% of test points certified on all 11 benchmarks, at both split fractions (0.5, 0.2). On the same splits the plug-in version of Table 8 still "certifies" 31% of NCI1 (fit 0.5) and 52% of NCI109 (fit 0.2), so the reviewer is right: the Table 8 rates are threshold-firing diagnostics, not validated coverage. The revision labels Table 8 as heuristic, adds the valid 0% column, and removes "covering most predictions on three of eleven benchmarks" from Section 4. The worst-case range term (N bars stacked on one landmark) enters at order 1/n and dominates the variance at these class sizes; the valid certificate fires on a synthetic population in our unit tests, so this is a limit of the bound, not a bug. To test whether it is a sample-size limit we repeated the valid certificate on Orbit100k's diagrams at class sizes m ∈ {1,000, 2,000, 5,000, 10,000, 20,000} (3 folds, seed 0, same split rule): valid coverage is 0% up to m = 5,000, 0.6% at 10,000 and 12.1% at 20,000 (84.7% accuracy on the certified points), while the median ratio of the certified gap to the required gap rises 0.002, 0.008, 0.042, 0.116, 0.353. The certificate becomes operational at about 20,000 modules per class; Orbit5k's 900 per class is 20× short. The revision reports this curve in place of Table 8's plug-in rates.

3. Two corrections we found while re-checking (not raised by reviewers). (a) The released PALACE code scales every embedding by a constant 2^{-3/2} that the paper's Definition 2.1 does not carry. Our audits compared code-scale distances with the paper-scale floor. Theorem 2.3 and all accuracies are unaffected; the tightness numbers change: the floor is 12.5× (min) / 29.8× (median) below realized distances on Orbit5k at α = 4 (paper: 4.4× / 10.5×), so "tight within a small factor" becomes "within one to one-and-a-half orders of magnitude", and Corollary 3.2's coverage drops to 0–0.5%. (b) The coherence audit tested the active set at r ≥ τ/4 instead of Definition 2.4's τ/(4ω_s); under the exact definition the α = 1.75 figure is 94.8% coherent (paper: 97.5%; realized-distance form 89.5%); α = 4 with the radius cap stays at 100%. Both corrections are in the revision with the audit scripts.

4. Significance (1Xuq Q2). Re-runs with identical folds, 3 seeds × 10 folds, Nadeau–Bengio corrected paired test: single-bifiltration COMPLEX (91.74) vs. ECS+XGB on the same bifiltration +2.50 (p = 2.7e-6), vs. HT +4.52 (p = 1e-9), vs. one-parameter PALACE +4.38 (p = 1e-10); the two- vs. one-parameter gain at α = 1.75 is +3.89 (30/30 folds, p = 7.7e-9). Against published numbers (Welch from reported mean ± s.d.) COMPLEX ties ECS (−0.06, p = 0.90), PALACE (+0.44) and Persformer (+0.54, p = 0.30) and beats Graphcode (+3.24, p < 0.01). Our in-protocol ECS re-run (89.2 over 3 seeds) is 2.6 points below the published 91.8, so the revision says "ties the best published results; beats ECS re-run on the same bifiltration", not "beats ECS".

5. What the paper claims after these changes. Deterministic two-sided bound (Prop. 2.2 + Thm 2.3) under a checkable coherence condition that holds on 95–100% of audited pairs and on 99.3% of 204,020 exhaustively checked pairs; a lower gauge that no prior multiparameter feature map has; a measured faithfulness that is real but local (the certificate protects a margin, not correctness, and the accurate head is the kernel SVM). The per-prediction probabilistic certificate is withdrawn as a claim. The two-sided bound, the exhaustive audit and the significance results are the contribution, and the paper now says only that.

AuthorsOpenReview ↗

Anonymized code release and variant map (Reviewer Xvs5 W2 / Q2a)

The anonymized code release promised in the global response is at

https://anonymous.4open.science/r/complex-anon-release-C220/

It contains every experiment and audit script behind the submission and the rebuttal (experiments/), the SLURM job files that ran them with the exact command line and task map in each header (experiments/cluster/), and the result files the tables, figures and audit paragraphs are read from, including the per-fold and per-seed records (experiments/results/; 610 files). The README carries the variant-mapping table the reviewer asked for: one row per table, figure and audit of the submission giving the script, the command or job task, weighted or unweighted stack, radius cap on or off, pooled or per-degree homology, bar budget NN, α\alpha, KK, net, seeds and folds, and the results file; a second table maps each rebuttal experiment (proof-valid certificate, class-size curve, exhaustive coherence, corrected audits, coherence-vs-accuracy, paired tests, D-COMPLEX re-run, definition unit tests) to its script, job and output. Nine cells are marked "unverified" where the exact producing command could not be recovered from any job file (for example the seed-2 per-fold file of the three-bifiltration run and the exact invocation behind Table 9); we flag these rather than guess. Author names, cluster account and host names, user ids and home directories are replaced by placeholders; nothing else was edited. The release depends on the PALACE code through experiments/pleb_bridge.py, whose comment block documents the 2−3/22^{-3/2} scale factor, the radius-cap switch and the farthest-point-sampling metric caveat that determines which PALACE checkout reproduces which run.

The revised PDF reproduces a compact version of the mapping table as an appendix and carries every change described in our replies (Table 8 relabelled as a plug-in diagnostic with the proof-valid column and the class-size curve, tightness ratios on the paper's scale, the exact-definition coherence audit, the exhaustive check, the min-vs-quantile wording, the rescoped Lipschitz claim, the single-bifiltration headline with paired significance tests, the qualified graph comparisons, the corrected D-COMPLEX table, the multiparameter contribution list, the coherence intuition paragraph and the settings-by-hypothesis table).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.