acceptodds
Under review as a conference paper at ICLR 2027

Higher-Order Persistence Diagrams

Abstract

Persistent homology tracks when topological features appear and disappear across filtered data and represents the lifetime of each feature by a persistence interval. In multiple instance learning, when the instances in a bag use filtrations over the same ordered parameter space, we can compare relationships between persistence intervals across instances. No existing method is designed to aggregate collections of persistence diagrams in multiple instance learning into a single persistence diagram of implications between intervals. We introduce higher-order persistence diagrams to represent relationships between persistence intervals and recursively construct each higher-order persistence interval from a pair of lower-order persistence intervals. We then aggregate these relationships across the persistence diagrams in a bag. Explicit construction of this aggregate is computationally expensive, so we introduce harmonic aggregation to compute coordinates of the aggregate from the lower-order persistence diagrams without explicitly constructing it. Computing a harmonic aggregation coordinate has lower asymptotic computational complexity than explicitly constructing the aggregate. We evaluate harmonic aggregation against persistence landscapes, persistence images, and PersLay on random network model and real-world graph-classification benchmarks, and find that harmonic aggregation has the highest accuracy on the random network model benchmark and on most real-world data benchmarks. Anonymized code is available at https://anonymous.4open.science/r/higher-order-PDs-BB69.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
persistent homology, multiple instance learning, topological data analysis, harmonic analysis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5YZ3OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Persistence diagrams as the fundamental tool of topological data analysis consist of a multiset of intervals. When summarizing a collection of such diagrams, the practitioners typically run them through vectorization techniques. They then usually use averages in the Euclidean or Hilbert spaces, and other tools to provide a common summary of a collection.

This paper adds additional structural information to this pipeline. It encodes the pairwise inclusions between the intervals in a diagram in the form of a higher order persistence diagram, and then, in principle, keeps repeating this iterative procedure to produce other higher order persistence diagrams. These are the additional derived combinatorial properties of the original persistence diagram. The paper then defines the harmonic aggregate as a tool to improve the explicit computation of some of the resulting coordinates.

In a summary, instead of vectorizing a persistence diagram, the paper proposes to derive some combinatorial properties of the diagram first and then vectorize the extended information. The paper states a stability result and reports on some experiments that indicate potential competitiveness in some settings.

Critical strengths
  1. The core idea of the paper is interesting, aggregating the combinatorics of inclusions of the intervals. Theorem 3.11 provides an interesting shortcut towards the computational goals.
Critical weaknesses
  1. The paper could lay out the overarching story and potential contribution more clearly, especially when it comes to the positioning of the results of the experiments. Essentially, instead of vectorizing the persistence diagrams, we first compute some derived parameters, and then vectorize the extended collection of parameters. As we are not introducing any new information, the contribution could be thought of just as an extended vectorization. Since we are starting with strictly more parameters, one would expect that the experimental results would never fall below the standard only-vectorization approaches. In this light the experimental results are not exactly spectacular. And crucially, it is not clear what the running times and memory requirements are compared to the previous methods, nor what the attributes in the use of the other vectorization techniques are. It seems plausible that by simply increasing the stated attributes of the other vectorizations performs better than the proposed method.

  2. Theoretical setup is vague and contains many errors and undeclared settings, the presentation is sparse on some crucial details, etc. The Stability Theorem 3.12 reads as very strong, but seems to come with many hidden caveats, so its strength is not really apparent.

Questions for authors

Questions:

In lines 86-78 it is stated that the same intervals have the same interpretation. Could you explain what is meant by that? In the real world datasets each interval has a representative and those are different. There is a paper [DOI: 10.4230/LIPIcs.SoCG.2026.94] that actually studies the relationship between the barcodes by comparing different intervals even topologically, not just at the level of isolated summaries. In this sense, the mentioned paper is strictly stronger. The same goes for lines 142-145.

Theorem 3.12: What are actually the assumptions here? If the space is supposed to be uniformly discrete (hence have a bounded number of intervals), then this does not imply stability when mapping from the pointsets (in the Gromov-Hausdorff metric) via the Vietoris-Rips or Cech filtrations to the summary, as the number of the intervals can increase. What is more, if uniformly discrete means that no intervals should be within ε\varepsilon, as lines 259-260 suggest, then the result is even weaker. This is a serious constraint and should be pointed out clearly.

Comments:

From 152 on it is not clear whether XX is a metric space or a space of intervals. Different notation is used. What is the order on the intervals?

Definition 3.2: It is not clear where the stated notions are defined. in 1., n≥1n \geq 1 should be n>1n>1. Later, pp is undefined.

Lines 184-187: Use of X/AX / A and X∖AX \setminus A.

Figure 2, Figure 3: The right part is incomprehensible.

Lines 218-219: What is a persistence diagram space?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 5YZ3 (1/2)

The review treats Vietoris-Rips and Cˇ\check{\mathrm{C}}ech filtrations of point clouds as the setting of our theory. Our setting is more general. A filtration of an abstract simplicial complex need not come from a point cloud, an embedding, or a real-valued threshold. Persistence diagrams are formal sums on metric pairs, of which the real birth-death half-plane with its diagonal is just one example. The 2022 paper Virtual persistence diagrams, signed measures, Wasserstein distances, and Banach spaces is one example of this broader literature. Many applications naturally give uniformly discrete spaces of persistence intervals, including cubical filtrations of digital mammograms by grayscale intensity and of lung tumor CT volumes by voxel Hounsfield values. Our benchmarks already demonstrate uniformly discrete filtrations on network datasets, from random network models to molecular graphs. The analytic consequences of uniform discreteness remain largely unstudied. Standard persistence diagram analyses almost never use the positive minimum separation between intervals, even when it holds naturally in the application domain. The observation that we take advantage of is that when the space of persistence intervals is uniformly discrete, the group of virtual persistence diagrams is discrete and locally compact in the topology given by the 1-Wasserstein distance. This lets us use Fourier analysis for our "Harmonic Aggregation". The theoretical results are stated for a broad class of metric spaces and preorders because different applications use different filtrations. The experiments use filtrations which give uniformly discrete metric spaces of persistence intervals. The review alleges "many errors and undeclared settings," but it does not identify an error in the stated theory. The second question asks for a theorem about point clouds, which are not the setting of this paper.

Question 1: The review states, "In lines 86-78 it is stated that the same intervals have the same interpretation." The point is when the features appear and disappear, not which cycles represent them. Equal intervals mean the features appear and disappear at the same filtration values. If one interval lies inside another, the first feature is present only when the second is also present. That is the implication we use to define higher-order persistence diagrams.

The review calls mixup barcodes "strictly stronger." Stronger than what? A mixup barcode describes what happens to homology classes under a specified inclusion of filtrations. We use the metric on persistence intervals to study the topology of the space of persistence diagrams and compute relations among intervals from the diagrams alone. The constructions have different inputs and answer different questions. The filtrations in a bag need not come with inclusions between them. If "strictly stronger" just means using information discarded by a persistence diagram, then keeping the original data is strictly stronger than computing persistent homology at all. That is not a comparison of the two methods, and it does not address what our construction recovers from persistence diagrams alone.

Question 2: Theorem 3.12 assumes that the order-one space of persistence intervals is uniformly discrete after collapsing the diagonal. The filtrations in our experiments satisfy this assumption.

Uniform discreteness does not bound the number of intervals. For example, let the filtration levels be the integers; then the birth-death pairs (b,d)(b,d) with b<db<d lie in Z2\mathbb{Z}^2. With the ℓ1\ell^1 metric, distinct pairs are at least 11 apart, and each pair is at least 11 from the diagonal. The quotient therefore has minimum distance 11 and is thus uniformly discrete. Even just the pairs (k,k+1)(k,k+1) give infinitely many possible persistence intervals. The claim that "if the space is supposed to be uniformly discrete, [then it has] a bounded number of intervals" is just false.

The assumption of uniform discreteness holds naturally in several applications. For example, an 8-bit grayscale cubical filtration has only 256 possible intensity levels, so there are even finitely many possible finite birth-death pairs across all images on that scale. Our network benchmarks also use discrete filtration levels, given by edge-entry ranks for random networks and trussness-based entry levels for TU graphs. The full real birth-death half-plane in the reviewer's point-cloud example is not uniformly discrete. The spaces of persistence intervals in our experiments are.

AuthorsOpenReview ↗

Response to Reviewer 5YZ3 (2/2)

Comment 1: The paper states, "Assume that (X,d)(X,d) is a metric space with a preorder ⪯\preceq," and then says, "We call an element of X(n)X^{(n)} an order-nn interval." These statements concern different levels of the construction. Item 2 of Definition 3.2 gives the order explicitly: (a,b)⪯(n+1)(c,d)(a,b)\preceq^{(n+1)}(c,d) if and only if c⪯(n)ac\preceq^{(n)}a and b⪯(n)db\preceq^{(n)}d.

Comment 2: The condition n≥1n\geq1 in item 1 is correct. The first clause defines X(1)X^{(1)}, and the recursive clause at n=1n=1 defines X(2)X^{(2)}. Changing the condition to n>1n>1 would leave X(2)X^{(2)} undefined. The paragraph immediately before Definition 3.2 defines pp: "Fix p∈[1,∞]p\in[1,\infty]."

Comment 3: The two notations have different meanings. X(n)∖A(n)X^{(n)}\setminus A^{(n)} removes the diagonal. X(n)/A(n)X^{(n)}/A^{(n)} collapses it to a single point. The first is the domain of the restricted metric in Lemma 3.4. The second retains the point used for matching intervals to the diagonal.

Comment 4: In the right panel of Figure 2, the two patterned regions represent order-two intervals that share one order-one interval. Their pair uses three distinct order-one intervals, as the caption states. The caption of Figure 3 identifies the pins as order-one persistence intervals, their heights as signed multiplicities, and the yellow surfaces as higher-order persistence intervals.

Comment 5: Immediately before Definition 3.3, the paper states, "Persistence diagrams are finite multisets of persistence intervals." Definition 3.3 then states, "For n≥1n\geq1, define the order-nn persistence diagram monoid by D(n)(X,A):=D(X(n),A(n))D^{(n)}(X,A):=D(X^{(n)},A^{(n)})." In the passage under discussion, the paper has already specified the corresponding virtual diagram space: "For each n≥1n\geq 1, let K(n)(X,A)K^{(n)}(X,A) denote the corresponding order-nn virtual persistence diagram group."

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

The paper defines a hierarchy of interval spaces and persistence-diagram monoids by recursively pairing lower-order intervals. It aggregates within-instance comparable interval pairs, with coefficients given by products of their multiplicities, before summing or averaging across a bag. Harmonic aggregation evaluates linear character-based coordinates of this aggregate. The main computational contribution is an exact factorization into weighted dominance sums, avoiding explicit enumeration of the quadratic pair aggregate at fixed order dimension. A Lipschitz bound is established under uniform discreteness. Experiments compare a mixture of first- and second-order HA features with persistence images, persistence landscapes, and PersLay using nested cross-validation on generated network-model bags and eight TU graph datasets grouped into same-class bags.

Critical strengths
  1. A concrete representation with an efficient evaluation route. Definition 3.7 records comparable interval co-occurrences within each instance before bag aggregation, and Theorem 3.11 gives an exact factorization of the resulting coordinates (pp. 5–7, 17). This connects an interpretable aggregate with a computationally practical alternative to explicit pair enumeration.
  2. Careful evaluation design in several important respects. The methods share input diagrams and downstream classifier protocols; transformation fitting and PersLay training stay within training partitions; real-world graphs are split before bags are formed (pp. 8–9, 13–15). The paper also reports paired uncertainty for its key synthetic ablations instead of relying only on point estimates.

The supplied primary benchmark also reproduced Table 1 in the October 7 reviewer-side run (HA .954, PI .940, PL .786, PersLay .889). This is positive numerical evidence for the main comparison, separately from the pending component-attribution and real-world checks.

Critical weaknesses
  1. The incremental value of the proposed relationship features is not established convincingly. Removing higher-order features changes synthetic accuracy by −0.008 with paired 95% CI [−0.019, 0.003], and removing the preorder changes it by −0.007 with CI [−0.017, 0.003] (p. 8). The first-order control is already capacity-matched at 256 complex/512 real features (p. 14); these intervals permit a benefit but do not resolve one. Table 2 lacks the corresponding real-world first-order/no-preorder controls and paired uncertainty, and evaluates artificial same-class bags at sizes 5–7 (pp. 9, 15). Separately, the paper's cross-method recipe gives HA 512 real features versus 72 for PL and 64 for PI/PersLay (p. 13); this is distinct from the capacity-matched internal control. Comparable capacity/tuning would aid attribution. The synthetic filtration also uses generator-provided probabilities or formation order. All methods receive the same diagrams, so this is not an unequal-input accusation; it limits what the task establishes about general-purpose graph bags.

  2. The released configuration does not yet identify the experiment behind the tables. The random-benchmark entry point selects eight landscapes at resolution 64, four PI bandwidth candidates and C from 10⁻³ through 10⁴; Appendix A specifies three landscapes at resolution 24, one bandwidth and C from 10⁻² through 10² (pp. 13–14). The October 7 reviewer-side primary benchmark, using 1,000 freshly generated batches / 4,000 graphs and unchanged author cross-validation and analysis functions, reproduced Table 1 to three decimals: HA .954, PI .940, PL .786 and PersLay .889, with corresponding macro-F1 also matching. A documented baseline-only metadata adaptation allowed primary analysis before the full-suite generation finished; it did not change scientific code, input graphs, or parameters. The README HA .941 is therefore stale documentation, not a demonstrated failure to reproduce. Torch is omitted from the dependency list and PersLay requires CUDA; Appendix A states the GPU requirement correctly. The exact configuration lineage remains a clarification request despite the positive main-table reproduction. Full ablation and real-world results were not complete at the October 7 evidence cutoff.

Questions for authors
  1. Exact experiment provenance: Which immutable source version, configuration and predictions produced Tables 1–2 and the ablations? Table 1 reproduced successfully with the supplied code. Please reconcile the landscape resolution/count, PI bandwidth and C-grid text with the released configuration, update the stale README and declare the Torch/CUDA requirements; state which results, if any, are affected.

  2. Incremental benefit: Do paired held-out real-world comparisons with the first-order and no-preorder controls establish a benefit from the new relationships under comparable feature capacity and tuning? Please report effect sizes, uncertainty and retained bag counts for available comparisons. The synthetic internal first-order control is already capacity-matched.

  3. Application scope: Which reported evidence survives a common filtration that does not use generator-specific probabilities or formation order, and less homogeneous bag composition? If such evidence is unavailable, please distinguish the present controlled same-class multiple-instance task from broader graph-bag applications. This is a scope/interpretation question, not a request to replace the mathematical construction.

2 comments

Thank you for response. The benefit of the higher-order features and the preorder on real data remains unclear to me. Do you have paired results on the same held-out TU graph bags comparing full HA with a capacity-matched first-order variant and a no-preorder variant? If so, please share the paired differences, confidence intervals, and bag counts.

AuthorsOpenReview ↗

Extending the Ablation to the Real World Datasets

Thank you for the careful review and for reproducing the main benchmark results. We hadn't run these ablations on the TU datasets before. We ran both variants for this response on the same held-out bags used in Table 2. Like full HA, the first-order-only variant uses 256 complex coordinates.

The table gives the accuracy differences in percentage points and also gives their paired 95% confidence intervals.

DatasetBag sizeHeld-out bagsFull HA minus first-order-onlyFull HA minus no-preorder
AIDS5400+0.25 [0.00, 0.75]0.00 [0.00, 0.00]
AIDS63300.00 [0.00, 0.00]0.00 [0.00, 0.00]
AIDS72800.00 [0.00, 0.00]0.00 [0.00, 0.00]
BZR5750.00 [0.00, 0.00]0.00 [0.00, 0.00]
BZR6610.00 [0.00, 0.00]0.00 [0.00, 0.00]
BZR7550.00 [0.00, 0.00]0.00 [0.00, 0.00]
COX25900.00 [0.00, 0.00]0.00 [0.00, 0.00]
COX26750.00 [0.00, 0.00]0.00 [0.00, 0.00]
COX27620.00 [0.00, 0.00]0.00 [0.00, 0.00]
MUTAG5350.00 [0.00, 0.00]0.00 [0.00, 0.00]
MUTAG6300.00 [0.00, 0.00]0.00 [0.00, 0.00]
MUTAG7200.00 [0.00, 0.00]0.00 [0.00, 0.00]
Mutagenicity5865-0.58 [-1.62, 0.35]0.00 [0.00, 0.00]
Mutagenicity6720-1.39 [-2.92, 0.00]0.00 [0.00, 0.00]
Mutagenicity7615-0.98 [-2.11, 0.16]0.00 [0.00, 0.00]
NCI15820-0.73 [-2.20, 0.73]0.00 [0.00, 0.00]
NCI166800.00 [-0.44, 0.44]0.00 [0.00, 0.00]
NCI17580-0.69 [-2.59, 1.21]0.00 [0.00, 0.00]
NCI10958230.00 [-0.85, 0.85]0.00 [0.00, 0.00]
NCI1096685-0.15 [-0.73, 0.29]0.00 [0.00, 0.00]
NCI1097585+0.68 [0.17, 1.37]0.00 [0.00, 0.00]
PROTEINS5220+4.55 [0.91, 8.18]+0.91 [-1.36, 3.18]
PROTEINS6185+5.41 [0.54, 10.27]+1.62 [-1.62, 4.86]
PROTEINS7153+3.92 [-0.65, 8.50]+1.31 [-2.61, 5.23]

Extending the ablation to the real world datasets, particularly for the PROTEINS dataset, seems to give particularly strong evidence for the higher order terms and statistically insignificant, but suggestive, evidence, for the preorder.

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

The paper addresses multiple instance learning where each instance is a persistence diagram computed over a shared filtration range. It introduces “higher-order” diagrams, built recursively from pairs of nested intervals (one interval containing another), and aggregates them over the bag by summing or averaging. “Harmonic aggregation” computes Fourier-type features of this aggregate without listing all pairs, reducing the cost from quadratic to near-linear in the number of intervals. A stability bound is proved for filtration values on a discrete grid. Bags are represented by averaged features and classified with an SVM. Overall, the paper adds a lot of new machinery for persistence diagrams but uses it only with an SVM on small graph benchmarks. Without showing the intended application, and how would HA work with current models such as GNNs or attention-based MIL.

Critical strengths

The paper offers careful formalization and a neat computational idea. The recursive hierarchy, including metrics and propagation of uniform discreteness across orders, is well defined, and the proofs I checked are correct. Factorization gives second-order Fourier coordinates of nested-interval pairs in O(s log s) rather than O(s²).

Critical weaknesses
  1. The contribution is limited. Bag-level HA is simply mean pooling of per-diagram features, as in the baselines, with interval pairs formed only within diagrams. First-order HA resembles standard random Fourier features (cf. PWGK), while the character framework comes from Fanning & Aktas (2026c). Theorem 3.11 reduces to algebraic factorization and classical dominance counting (Bentley, 1980). Theorem 3.12 provides a generic stability bound that deteriorates as the grid is refined. Crucially, higher-order features show no significant benefit, because replacing them with first-order features changes accuracy by only 0.008, and orders ≥3 are untested.

  2. The experiments validate neither MIL nor higher-order structure. Bags contain same-class graphs, making this set classification rather than genuine MIL. Baselines are weak and dimension-imbalanced (64–72 vs. 512 features), excluding MIL methods, diagram kernels and graph models.

  3. Interpretability is never evaluated.

Questions for authors
  1. Can you compare with WL/GIN bag voting, DeepSets or attention MIL, PD kernels, and dimension-matched or tuned PI/PL?
  2. How does HA perform on a genuine MIL task, for example the microscopy and whole-slide image datasets used by TopoMIL and PMIL, compared with these methods and with attention-based MIL (Ilse et al., 2018)?
  3. How does HA compare with standard graph classifiers such as the WL kernel and GIN on the TU datasets, both on single graphs and on bags (with per-graph predictions averaged or voted)?

Comments on the paper

1 comment

Dear Reviewers,

As the discussion period draws to a close, please read and acknowledge the authors’ rebuttal within the next 24 hours. I also encourage you to engage in the discussion and share any remaining questions or concerns.

Thank you for your time and thoughtful reviews.

Best,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.