{"id":"497ac641-26cc-49d3-aefe-ceb90df2a99d","arxiv_id":"2510.12756","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Averaging representative cycles over perturbed filtrations yields persistence heatmaps that are Lipschitz-stable and localize SVM-learned feature importance back into data space.","lead":"Scientists studying point-cloud shapes (loops, holes) get unstable cycle representatives when the data wiggles; this paper averages them over random perturbations to produce a 'persistence heatmap' that is provably Lipschitz-stable, and lets an SVM pick which topological features to emphasize for a given classification or regression task. The payoff: a principled way to visualize where a topological ML model's signal actually lives in the data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Definition 3.2's convolution integrates η off its domain C; as stated, η∗K is undefined for Gaussian (and boundary-near compact) kernels, so Theorems 2–4 lack a well-defined object.","rationale":"The reader's verdict is CONDITIONAL, and I largely agree. I independently re-derived the convolution bounds in Theorem 2 and Cor 4.3; for a globally defined bounded η they are standard. The weakest point, however, is earlier than Cor 5.5–5.7: Definition 3.2's η∗K is not well-defined. The persistence heatmap is introduced as a map from the monotone cone C⊂R^k to R^k (§3). Convolving it against a kernel requires evaluating η at x−ε for ε outside the cone. A Gaussian kernel has positive mass outside C for every x; compact kernels also fail near ∂C. Nothing in §4 specifies an extension. So Theorems 2–4 are theorems about a partially defined expression. This is a correctness risk at the level of the paper's main stability claim, not just the visualization layer. The reader's cited Π-Lipschitz issue is real but secondary. Both are patchable: one can extend η or restrict α; one can prove a grid-dependent Lipschitz bound for Π. Because these fixes do not require changing the overall approach, I would not reject; the conditional verdict stands.","tokens_in":15317,"tokens_out":9772,"duration_ms":83392,"concrete_test":"Take the minimal simplicial complex with vertices a,b and edge ab, so C={x∈R^3: x1≤x3, x2≤x3}. Let x=(0,0,1) (interior) and K be the Gaussian kernel from Cor 4.3. Compute the set {ε: x−ε∉C}; its Gaussian measure is >0. Then define two extensions of η to R^3: η≡0 outside C and η≡c (a nonzero constant) outside C, and compute η∗K(x) under each. If the two values differ, Def 3.2 has no canonical meaning without an extension, so Theorem 2/Cor 4.3 cannot be evaluated as written. For a compact-kernel version, choose α>dist(x,∂C) and compare the same two extensions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Definition 3.2 defines the expected heatmap as η∗K(x)=∫ η(x−ε)K(ε)dε, with η:C→R^k and C the monotone cone. The paper never extends η to R^k. For any Gaussian kernel (Cor 4.3), supp(K)=R^k, so for every x∈C the set {ε: x−ε∉C} has positive measure and η(x−ε) is undefined there; the convolution integral is not well-defined. For compactly supported kernels (Cor 4.1/4.2), the same failure occurs whenever B_α(x)⊄C, which is not excluded by the hypotheses. The proof of Theorem 2 integrates η over all of R^k; this presupposes an extension whose choice changes the summary and is not specified. Thus the central stability theorems—the paper's main theoretical contribution—are currently statements about an object that is not defined. This is distinct from the reader's secondary concern about Π (Cor 5.5–5.7), which affects only the visualization layer; the domain problem invalidates the core η-level stability claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces persistence heatmaps, a weighted-sum summary built from representative cycles, birth/death simplices, and bounding chains in annotated persistence diagrams, and proposes an “expected persistence heatmap” obtained by convolving the heatmap with a kernel. It claims uniform continuity (Theorem 1) and Lipschitz stability with explicit constants for triangular, Epanechnikov, and Gaussian kernels (Theorems 2–4, Corollaries 4.1–4.3). It further defines a visualization pipeline that maps a heatmap to a voxel grid, claims corresponding stability for the visualized heatmap (Theorem 5, Corollaries 5.5–5.7), and demonstrates SVM-based task-specific heatmaps on annular point clouds, Ginibre-ensemble eigenvalues, and linked twist maps.","tokens_in":15530,"tokens_out":6251,"duration_ms":52435,"significance":"If valid, the proposed persistence heatmap would give a stable, explainable TDA summary that localizes topological features in the original data space, addressing the known instability of representative cycles. The paper's theoretical framework is natural and, modulo the gaps below, the convolution-based stability arguments are standard and potentially useful. The computational examples showcase an interesting application of task-specific summaries. However, the core theoretical claims are currently undermined by an undefined convolution and an unsupported Lipschitz assumption in the visualization layer; these issues must be resolved before the paper's central promises can be accepted.","major_comments":[{"comment":"The expected persistence heatmap η∗K(x)=∫η(x−ε)K(ε)dε integrates over all R^k, but η is only defined on the monotone cone C⊂R^k. For the Gaussian kernel (Cor 4.3), supp(K)=R^k, so for every x the set {ε: x−ε∉C} has positive measure and η(x−ε) is undefined there. For compact kernels (Cor 4.1–4.2), the same issue occurs whenever B_α(x)⊄C, which is not excluded. The proofs of Theorems 1–4 integrate η over R^k, implicitly requiring an extension. Since the choice of extension alters the summary and the stability constants, the central stability results are currently statements about an undefined object. Please specify an explicit extension of η to R^k (e.g., zero, constant, or nearest-point extension) and re-state and prove the theorems under that choice.","section":"Definition 3.2, Theorems 1–4, Corollaries 4.1–4.3"},{"comment":"Even if an extension is supplied, the local hypothesis “∥η∥_∞<M on B_{2α}(x)” only makes sense if B_{2α}(x) is contained in the domain on which η is actually defined. If η is defined only on C, this requires an explicit assumption that B_{2α}(x)⊆C; otherwise the convolution integral on the right-hand side is not defined. The proof uses values of η on B_α(u)∪B_α(v), which lie in C only if B_{2α}(x)⊆C. This condition is missing from the statement and from the corollaries.","section":"Theorem 3 and Corollaries 4.1–4.2"},{"comment":"The claimed Lipschitz constants for Θ∗K depend only on M, k+nd, and α, and omit any dependence on grid resolution or simplex geometry. This requires the voxel-projection map Π of Definition 5.3 to be Lipschitz in the geometric-realization coordinate y with a uniform constant. The paper neither states nor proves such a bound. Π is a piecewise volume-ratio map: for a simplex σ, Π_i,φ(σ) is the fraction of σ's volume in voxel U_i. As φ varies, this fraction can change at a rate that blows up when the simplex volume tends to zero (e.g., for a segment of length L, the fraction in a voxel changes at rate ~1/L) or when a vertex crosses a voxel boundary. Thus the uniform Lipschitz claim is unsupported. The authors should add explicit geometric assumptions (e.g., a lower bound on simplex volumes and a bound on aspect ratios) and make the constants depend on them, or restrict the stability stateme","section":"Definition 5.3, Theorem 5, Corollaries 5.5–5.7"},{"comment":"The computational examples compute “expected PHMs” by averaging over 100 independently generated point clouds, but no kernel K or noise distribution is specified. The averaged object is therefore not the η∗K of Definition 3.6, and the experiments do not actually instantiate or validate the theoretical stability results. Please specify the sampling procedure and its connection to the convolution definition, or explicitly present the averaging as a heuristic approximation rather than a computation of the stable summary defined earlier.","section":"Section 6 (Computational Examples)"}],"minor_comments":[{"comment":"“Epanechinikov” should be “Epanechnikov”.","section":"Section 4.2"},{"comment":"The text refers to “lemma 5.2” and “lemma 5.3” but these are Definitions 5.2 and 5.3; please correct the cross-references.","section":"Section 5"},{"comment":"The paragraph describing class B vs class C says “visualized in fig. 7” twice; the second should refer to fig. 8.","section":"Section 6.1.3"},{"comment":"In the formula F(b_i, d_i)=v_i f_i + b_i/n, the symbol b_i overloads the death value and the SVM bias b; please rename to avoid confusion.","section":"Section 6 (Method)"},{"comment":"Reference [12] appears to be about DNA microarrays, which does not match the linked twist map citation; please verify and correct.","section":"References"},{"comment":"The kernel is introduced as K:R^k→R^k and later as component functions K_i:R^k→R; the normalization ∫K(ε)dε=1 should be clarified as a component-wise condition.","section":"Definition 3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper’s self-citation to [2] is disclosed and the convolution inequalities are standard; the novelty lies in the persistence-heatmap construction and its application. However, the undefined convolution in Definition 3.2 is a fundamental gap that invalidates the main theorems as written, and the visualization-layer Lipschitz claim in Corollaries 5.5–5.7 appears to require additional geometric hypotheses. Both are fixable in principle, but they require substantial rewriting of the theoretical sections. I would encourage the authors to address these points carefully before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my honest read. The core idea — average unstable representative cycles over perturbed filtrations to get a stable heatmap — is a good one and directly addresses a real bottleneck in TDA-for-ML. The persistence heatmap (Def 3.2), the SVM-parametrized expected version (Sec 6), and the voxel-projection visualization (Defs 5.2–5.3) are genuinely new objects, and the computational examples give a fair sense of how they could be used. The paper is also upfront that Theorems 2–4 use the authors' prior stabilization result [2], and the Lipschitz constants for triangular, Epanechnikov, and Gaussian kernels do check out once the convolution is well-defined.\n\nBut the stress-test note is right, and it is not a minor quibble. Definition 3.6 defines η∗K(x) = ∫ η(x−ε)K(ε)dε, yet η is only defined on the monotone cone C. For a Gaussian kernel the support is all of R^k, so for every x in C the integrand is undefined on a set of positive measure. For compactly supported kernels the same failure occurs near the boundary of C, and the hypotheses of Theorems 2–4 do not exclude it. As written, the paper's central stability results are statements about an object that does not exist. This is patchable — extend η outside C in some explicit way and restrict the kernel support or the neighborhood so the convolution stays inside C — but the choice of extension changes the summary and the constants, so it has to be confronted directly.\n\nThe secondary issues are real but smaller. Corollaries 5.5–5.7, which carry the 'stable heatmaps on images' promise, are asserted without proof; the Lipschitz constant for the projection Π could depend on grid resolution and simplex geometry, and the paper gives no bound. The abstract's claim that summing pixel values gives the learned regression function also overstates what Section 5 shows — the sum is conserved, but it isn't obviously the SVM output.\n\nNet: the paper has a useful, well-motivated toolkit, but the main theorem needs a substantial correction before it is citable. A serious referee should engage with it because the fix is manageable and the intended application is valuable. Send it to peer review, but expect a major revision.","headline":"Good idea, clean exposition, but the main stability theorems are stated for an undefined convolution — fix the domain issue and this becomes a credible TDA explainability tool.","tokens_in":765,"tokens_out":845,"would_cite":false,"duration_ms":30146,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["55N31","62R40"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that averaging unstable representative cycles over random perturbations yields stable, class-revealing heatmaps—the expected persistence heatmap—with explicit Lipschitz bounds.","keywords":["persistence heatmap","representative cycles","persistent homology","topological data analysis","kernel stability","Lipschitz stability","explainable machine learning","annotated persistence diagrams"],"falsifier":"Compute the heatmap image for a single edge of weight 1 whose geometric realization crosses a voxel boundary, and move one endpoint by ε so the crossing point shifts. The fraction of the edge assigned to each voxel changes with slope that depends on the edge length and grid size, and if the edge is short relative to a voxel this slope can be arbitrarily large; repeating at two grid resolutions will show the Lipschitz constant is not independent of resolution, contradicting Corollaries 5.5–5.7 as stated.","tokens_in":15070,"feed_emoji":"🌡️","tokens_out":8259,"duration_ms":67298,"temperature":0.7,"pith_summary":"The central claim is that the unstable chains that come out of persistence algorithms—representative cycles, bounding chains, birth and death simplices—can be turned into stable, task-relevant visualizations. The paper defines a persistence heatmap that spreads each topological feature's weight over the simplices carrying it, and then averages these heatmaps over small random perturbations of the input. The main theoretical results are that this expected heatmap is uniformly continuous and Lipschitz-stable with explicit constants for triangular, Epanechnikov, and Gaussian kernels. The paper further shows how to weight the heatmap with SVM feature coefficients so that the resulting images localize the regions that drive a classification or regression. If the claims hold, this gives an explainability layer for TDA that preserves where a feature lives while remaining stable under noise.","feed_headline":"Averaging makes unstable topological cycles into stable heatmaps","feed_subtitle":"Averaging over perturbed inputs yields provably stable heatmaps showing where each topological feature lives in data.","key_machinery":"The central objects are the annotated persistence diagram—each birth–death pair decorated with the chain (representative cycle, bounding chain, birth simplex, or death simplex) that the persistence algorithm outputs—and the persistence heatmap η that distributes each feature's weight equally over the simplices in its chain. The stabilizing mechanism is convolution: η∗K averages these chain-weights over kernel perturbations of the input, and the key inequality is the kernel's L1 displacement bound ∫|Ki(s+t)−Ki(t)|ds ≤ b_i‖t‖, which converts a small input move into a small change in the averaged output. The same convolution is applied to the composed map Θ=Π∘(η×I) that sends weights and a geom","core_discovery":"On the paper's own terms, the discovery is that instability is not a property of the chains themselves but of choosing one chain among many equally valid ones; averaging over many perturbed inputs replaces the discrete choice with real-valued weights that vary smoothly. Concretely, for a persistence heatmap η built from the annotated chains and a kernel K with an L1 displacement bound, the expected heatmap η∗K is Lipschitz, with constants given for the three standard kernels. The same convolution argument is then applied to the composed map that projects simplices onto an image grid, yielding the claimed stable heatmaps on images whose pixel sums recover learned function values.","pith_inferences":["The stability of the image heatmap is not actually a corollary of the chain-stability theorems: it needs a Lipschitz bound on the voxel projection Π, which the paper leaves unstated; a grid-dependent constant would change the practical guarantee.","The same averaging recipe applies to any unstable function of the persistence diagram that is bounded and chain-valued—for example, representative cocycles or minimal generators should stabilize identically.","Because the heatmap is a real-valued vector indexed by simplices, it can be fed into any linear model, not just SVM, so the explainability layer generalizes to logistic or ridge regression with no change to the stability proof."],"forward_implications":["Representative cycles can be replaced by real-valued chain weights that change at most linearly with input perturbation, so localization maps become reliable across noise.","Choosing the kernel radius α gives an explicit sensitivity knob: with a triangular kernel, the heatmap is 2M(k+1)√k/α-Lipschitz, so larger α smooths more.","The visualization pipeline preserves total weight: summing pixel values recovers the total heat, which is what lets a learned regression value be reconstructed from the image.","The construction is algorithm-agnostic: any implementation that outputs annotated persistence diagrams can feed the same stabilization.","SVM feature coefficients can be folded into the heatmap, so the same stable image that localizes features also explains which regions push a classification one way or the other."],"fun_headline_variants":["Stable heatmaps from unstable topological cycles via averaging","Explainable TDA: averaging unstable cycles yields stable heatmaps","Average perturbed inputs to stabilize topological explanations","Unstable cycles become stable heatmaps through averaging","Provably stable heatmaps from averaged topological cycles"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The image-stability claim rests on the unproven premise that the voxel-projection map Π, which assigns each simplex's weight to grid cells by length or area fraction, is uniformly Lipschitz in the vertex coordinates; this can fail when a vertex crosses a voxel boundary or a simplex becomes degenerate, and the stated constants ignore grid resolution.","fun_headline_variants_meta":{"raw":{"variants":["Stable heatmaps from unstable topological cycles via averaging","Explainable TDA: averaging unstable cycles yields stable heatmaps","Average perturbed inputs to stabilize topological explanations","Unstable cycles become stable heatmaps through averaging","Provably stable heatmaps from averaged topological cycles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000173,"raw_usage":{"total_tokens":1038,"prompt_tokens":588,"completion_tokens":450,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":332,"completion_tokens_details":{"reasoning_tokens":376}},"tokens_in":332,"tokens_out":450,"duration_ms":4236,"temperature":1.0,"reasoning_tokens":376,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T09:52:33.685790+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the heatmap image for a single edge of weight 1 whose geometric realization crosses a voxel boundary, and move one endpoint by ε so the crossing point shifts. The fraction of the edge assigned to each voxel changes with slope that depends on the edge length and grid size, and if the edge is short relative to a voxel this slope can be arbitrarily large; repeating at two grid resolutions will show the Lipschitz constant is not independent of resolution, contradicting Corollaries 5.5–5.7 as stated.","supporting_citations":[],"review_version":1}