{"id":"f25ed161-c2e6-4e97-8275-c0251555ad38","arxiv_id":"2507.22816","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A finite sample of directions and filtration values of the persistent homology transform can be extended to the whole sphere with interleaving-distance error at most twice the sampling radius plus the height sampling error.","lead":"This paper shows that the persistent homology transform (PHT), a shape descriptor that records topological features from every viewing direction, can be interpolated from a finite sample of directions using Kan extensions, with a proven error bound. The result gives the first approximation guarantees for the PHT from finite directional and scalar data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Left-extension formula omits the special top morphisms at β=supX; under the displayed L^- definition the main error bound is false.","rationale":"The reader's weakest assumption concerned the practical need for inter-direction homology maps (Remark 3.11). That is a genuine limitation, but it is explicitly acknowledged. My stress-test found a more elementary issue inside the idealized module-level argument: the displayed past-light-cone set and the formally stated comma category disagree at the top of the filtration, and the main theorem is false for the displayed version. The counterexample is a single point, so it is easy to verify analytically. Since the paper also gives the comma-category definition and the theorem likely survives when that definition is used, the appropriate disposition is still conditional revision rather than rejection. Agreement with the reader is only partial: both concerns bear on the Kan-extension formalism, but the reader's stated weakest assumption does not identify the L^-/comma-slice discrepancy.","tokens_in":21090,"tokens_out":24528,"duration_ms":321823,"concrete_test":"Compute the left Kan extension for the one-point example M={p=(1,0)}, A={v=(1,0)}, X=I at query w=(cosδ,sinδ), once using the displayed L^- inequality and once using the under-comma category of A_X→S_X including the special morphisms. If the first module is identically zero while the true PHT_w is nonzero on [cosδ,1], the displayed definition is incompatible with Theorem 4.6; use the comma-slice version and re-check the construction of Ψ in Lemma 4.1 at β=1-d(v,w).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 defines A_X with an extra morphism (v,α)→(w,supX) for all v,w,α, so the under-comma slice ι↓(w,β) used in Theorem A.3 contains all objects (v,α) when β=supX. However, the paper then defines L^-(w,β) by the inequality α+d(v,w)≤β and calls it the projection of that slice; the two sets differ at β=supX. Lemma 4.1 and Theorem 4.6 are stated and proved for the L^- colimit, so this is not only a notation issue. Counterexample: take M={p=(1,0)}, A={v=(1,0)}, X=I=[-1,1], and query w=(cosδ,sinδ). For the displayed L^-, C_w(β)=colim over α+d(v,w)≤β of H_0(M_{v,α}); since H_0(M_{v,α})=0 until α=1 but α≤β-d≤1-d<1, C_w is identically zero. The true PHT_w has a nonzero bar from cosδ to 1, so d_I is infinite and Theorem 4.6 fails. If instead the colimit is taken over the full comma slice, including the special morphism to (w,1), C_w is supported at β=1 and the interleaving distance is 1-cosδ, which is finite and below 2d(v,w). The proof of Lemma 4.1 also silently uses the special morphism exactly when β+d(v,w)=supX; with the displayed formula the map Σ can fail. The fix is to define L^- as the actual comma slice, or explicitly add the special top objects, and then re-verify the examples.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method to interpolate the persistent homology transform (PHT) from finitely many sampled directions by viewing the direction-indexed PHT as a coherent functor on a space-time category and applying the Kan-extension theory of Bubenik, de Silva, and Nanda. Section 3 constructs left, right, and center Kan extensions in both Euclidean and angular coordinates, and Section 4 proves pointwise and global interleaving-distance bounds, including a fully discrete bound when both directions and filtration values are sampled. The stated main result, Theorem 4.6, is that the interleaving distance between the true PHT and its Kan-extension approximation is at most twice the Hausdorff distance from the sample direction set to the sphere, for every homological degree.","tokens_in":21450,"tokens_out":12671,"duration_ms":135975,"significance":"If correct, this is a genuinely useful interpolation theorem: it turns a finite sample of the PHT at the module level into a Lipschitz approximation with explicit, parameter-free error control, and it appears to be the first such approximation guarantee for the PHT from finite directional data. The derivation is parameter-free, the main machinery is credited to prior work by Bubenik, de Silva, and Nanda, and the paper honestly records the key practical limitation in Remark 3.11, namely that computing the extension requires inter-direction homology maps, not merely per-direction persistence diagrams. The examples are instructive, and the distinction between Euclidean and angular coordinates is thoughtfully developed.","major_comments":[{"comment":"The displayed definition of the past light cone, L^-(w,β) = {(v,α)∈A_X | α+d(v,w)≤β}, is not the projection of the under-comma category ι↓(w,β) when β = sup X, because A_X contains the always-morphism (v,α)→(w,sup X) for every (v,α). At β = sup X the comma slice contains all objects, while the displayed inequality omits those with α+d(v,w)>sup X, notably α = sup X whenever d(v,w)>0. This is load-bearing: Lemma 4.1 constructs the map Σ_{v,w} by placing (v,α) in the diagram that defines C_w(T_{d(v,w)}(α)), and when T_{d(v,w)}(α)=sup X this requires exactly the omitted special morphism. Concretely, take M={p=(1,0)}, A={v=(1,0)}, X=[-1,1], and w=(cos δ,sin δ). Under the displayed formula, C_w(β)=0 for all β, because the only nonzero H_0(M_{v,α}) occurs at α=1, which never satisfies 1+d(v,w)≤β; the true PHT has a bar [cos δ,1), so the interleaving distance is infinite and Theorem 4.6 is false for the extension as defined by the displayed set. The fix is to define L^- as the projection of the actual comma slice, or to add the special top morphisms explicitly, and then re-verify the proof and the examples; with that corrected definition the argument in Lemma 4.1 appears to go through, but the current text is internally inconsistent.","section":"§3.2, Lemma 4.1, Theorem 4.6"},{"comment":"Remark 4.2 restricts the proof to the left Kan extension, and the proof of Lemma 4.1 only constructs the interleaving for that case. Theorem 4.6, however, is stated for \"the Kan extension approximation\" without qualification, Section 3.3 presents the right and center extensions as equally valid alternatives, and Example 4.3 asserts the same bound for \"any of the three Kan extensions.\" No argument is supplied for the right or center extensions; duality is not automatic under the capped shift, and the center extension is an image rather than a limit or colimit. The theorem should either be explicitly restricted to the left Kan extension or supplemented with proofs for the other two.","section":"Theorem 4.6 and §4.1–4.2"},{"comment":"The right- and center-extension computations use the future light cone L^+(w,β) = {(v,α)∈A_X | β+d(v,w)≤α}, which is subject to the same top-parameter truncation as L^- once the always-morphisms to sup X are included in A_X. When β+d(v,w)≤sup X fails, the formal over-comma category still contains (v,sup X) by the special morphism, so the displayed L^+ omits objects that the categorical definition includes. The barcodes in Example 3.9 and Figure 4 should therefore be recomputed against the formal comma-category definitions, or the definition of L^+ should be corrected consistently. This matters because the paper never proves the error bound for these extensions, so the examples are the only evidence that they behave as claimed.","section":"§3.3, Example 3.9"}],"minor_comments":[{"comment":"The heading \"Lipshitz Stability\" and the word \"Lipshitz\" in Corollary 3.4 should be corrected to \"Lipschitz.\"","section":"§2.6 and Corollary 3.4"},{"comment":"The text \"described aboove\" contains a typo and should read \"described above.\"","section":"Theorem 3.10"},{"comment":"In the definition of A_X, the always-morphism is declared for any v,w∈S^{d-1}, but the objects of A_X are pairs (v,α) with v∈A; the declaration should be for v,w∈A, or the category should be defined on all of S^{d-1} and then restricted.","section":"§3.1"},{"comment":"The proof uses a one-sided notion of an ε_T-net, requiring that for every y there exists t'∈T with y≤t'≤y+ε_T, whereas the theorem statement says only that T is an ε_T-net of X. This stronger condition should be stated in the theorem, since a standard symmetric net near sup X need not provide such a t'.","section":"§4.4, Theorem 4.12"},{"comment":"This example is the tightness example and is correct in spirit, but after the correction of the L^- definition it should be re-annotated to show which step relies on the special morphism to sup X in the space-time category.","section":"Example 4.5"}],"recommendation":"major_revision","confidential_remarks":"The core idea is sound and the paper is well within the scope of the journal. The main obstruction is the inconsistency between the displayed past-light-cone formula and the formal comma-category definition; this is not a cosmetic issue, since it invalidates the proof of Lemma 4.1 and Theorem 4.6 as written. I believe the gap is fixable by defining L^- as the actual comma slice and re-verifying the subsequent arguments, and I therefore recommend major revision rather than rejection. The absence of a proof for the right and center extensions should also be addressed, either by restricting the theorem or by extending the argument."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's the take on 2507.22816. The paper's central result, Theorem 4.6, is false as stated. The displayed definition of L^-(w,β) in Section 3.2 omits the special top morphisms (v,α)→(w,sup X) that were added to the space-time category in Section 3.1. The under-comma slice at β=sup X contains all sampled objects, not just those satisfying α+d(v,w)≤β. The paper explicitly says L^- is the projection of that slice, but the formula is a strict subset at the top. The stress-test counterexample is correct: take M={p=(1,0)}, A={v=(1,0)}, X=I, and query direction w at small angle δ. The displayed L^- gives the zero module everywhere, while the true PHT has a bar [cosδ,1). The interleaving distance is infinite, so the 2·d_H bound fails. This is not a notation quibble; Lemma 4.1 and Theorem 4.6 are proved for the truncated colimit, and the proof's construction of Ψ silently relies on the special morphism exactly at β+d(v,w)=sup X. Theorem 4.12 inherits the same problem.\n\nThe fix is simple: define L^- as the actual comma slice, or explicitly add the top objects to the colimit. The stress-test notes that with the corrected definition, the counterexample gives a finite interleaving distance (1-cosδ) below 2d(v,w), so the theorem is likely salvageable. That is the key referee demand.\n\nCredit where due: this is the first application of the Bubenik–de Silva–Nanda interpolation machinery to the PHT, and the new angular coordinate system (Definition 2.7) is a real contribution. The paper is clearly written and the categorical setup is appropriate. The Lipschitz extension theorem (Theorem 3.10) is a nice conceptual step. The idea that module-level sample data, including inter-direction maps, can yield provable interpolation is genuinely useful for TDA shape analysis.\n\nSofter spots: the right and center Kan extensions are not covered by any error bound, and the examples for them look under-verified; the reader is right to flag possible inconsistencies with the formal definitions. The practical barrier of needing inter-direction homology maps is acknowledged in Remark 3.11 but remains a serious limitation.\n\nBottom line: this paper is worth a serious referee, but the current version cannot be accepted as is. The main theorem must be corrected and re-proved. If the fix goes through, the result is a solid contribution to the PHT literature.","headline":"Main theorem false as stated because the left-Kan colimit omits the special top-parallel morphisms; the gap is concrete and fixable, and the paper deserves a careful revision.","tokens_in":21953,"tokens_out":7122,"would_cite":false,"duration_ms":77301,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["55N31","18A40"],"pacs":[],"model":"deepseek-v4-flash","headline":"The persistent homology transform can be interpolated from finitely many sampled directions, with an explicitly bounded error.","keywords":["persistent homology transform","Kan extensions","interleaving distance","shape interpolation","topological data analysis","Lipschitz stability","o-minimal structures"],"falsifier":"Compute the left Kan extension for the single-point shape $M = \\{(0, 1)\\}$ with one sampled direction $v = (1, 0)$, query the south pole $w = (0, -1)$, and compare the birth positions of the true and extended $\\Theta$-persistence modules: the paper's Example 4.5 predicts an interleaving distance of exactly $\\pi$, matching the bound $2 d_g(v, w)$. More generally, a decisive test would evaluate, for a range of constructible shapes inside the unit disk and direction sets $A$, the true interleaving distance between $\\mathrm{PHT}^X_n(M)$ and its center Kan extension and check whether it ever exceeds $2 d_H(A, \\mathbb{S}^{d-1})$ (or $2\\varepsilon_A + \\varepsilon_T$ for fully discrete data).","tokens_in":20882,"feed_emoji":"📐","tokens_out":8105,"duration_ms":87317,"temperature":0.7,"pith_summary":"This paper asks whether the persistent homology transform (PHT) of a shape can be interpolated from finitely many sampled directions and whether the interpolation error can be bounded. The authors answer both questions affirmatively by working at the module level, where each sampled direction contributes not just a persistence diagram but the homology maps that connect sublevel sets across directions. Adapting the Kan-extension interpolation theory of [24,25], they construct left, right, and center extensions of the PHT and prove these extensions are 1-Lipschitz. Their main approximation results are Theorem 4.6, bounding the interleaving distance between the true PHT and the extension by twice the Hausdorff distance from the sampled direction set to the sphere, and Theorem 4.12, bounding the fully discrete case by $2\\varepsilon_A + \\varepsilon_T$. A correct proof would give the first provable scheme for approximating the PHT from finite data, with the error shrinking linearly as the directional sampling becomes denser.","feed_headline":"Direction sampling radius bounds the PHT interpolation error","feed_subtitle":"Kan extensions interpolate the transform across directions, with error at most twice the sample spacing.","key_machinery":"The load-bearing object is the space-time category $A_X$. Its objects are pairs $(v, \\alpha)$ consisting of a direction $v$ in the sample set $A$ and a filtration value $\\alpha$ in an interval $X$, and it has a morphism $(v, \\alpha) \\to (w, \\beta)$ whenever $\\alpha + d(v, w) \\le \\beta$, together with a top morphism to $(w, \\sup X)$ for every pair; Lemma 2.12 and Lemma 2.13 turn these inequalities into inclusions of sublevel sets, so the PHT data assemble into a functor $G : A_X \\to \\mathbf{Vect}$. The left, right, and center Kan extensions of $G$ along the inclusion $A_X \\to \\mathbb{S}^{d-1}_X$ then provide the interpolation formulas: colimit over the past light cone, limit over the future light cone, and the image of the natural map from the former to the latter. The coherence condition of [24,25] is exactly the requirement that the data form such a functor, and Proposition 3.3 proves that every PHT satisfies it automatically.","core_discovery":"The paper's central claim is that the persistent homology transform of a constructible set $M$ inside the unit disk can be interpolated from finitely many sampled directions, and the interpolation error can be controlled, provided the samples include the maps that homology induces between sublevel sets in different directions. The authors model the sampled data as a functor $G : A_X \\to \\mathbf{Vect}$ from a space-time pre-order category $A_X$ to vector spaces, and they construct three canonical extensions of $G$ to the whole sphere of directions: the left Kan extension (a colimit over past light cones), the right Kan extension (a limit over future light cones), and the center Kan extension (the image of the natural map between them). Theorem 4.6 states that for every direction and every homological degree, the interleaving distance between the true transform and any of these extensions is at most twice the Hausdorff distance from the sample direction set $A$ to the sphere. Theorem 4.12 extends the bound to fully discrete data, giving $2\\varepsilon_A + \\varepsilon_T$ when directions form an $\\varepsilon_A$-net and filtration values form an $\\varepsilon_T$-net. The example of the unit disk shows that the left and right extensions can introduce spurious bars where the center extension does not, and Example 4.5 shows the factor of two in the pointwise bound is sharp.","pith_inferences":["A practical implementation of this scheme would need to estimate the inter-direction homology maps from persistence diagrams via coherent matchings; the paper explicitly leaves that estimation to future work, so it is the main bottleneck between the theorem and a software pipeline.","The same space-time category construction should transfer to the Euler characteristic transform and Betti curve transforms, since the only geometric input needed is the nesting of sublevel sets; that would give analogous interpolation bounds for those signatures.","Lemma 4.10's use of angular coordinates hints that angular sampling grids, whose spacings are measured geodesically, yield strictly better constants than Euclidean grids of the same cardinality; a direct experimental comparison of the two sampling schemes on non-convex shapes would settle which is preferable in practice."],"forward_implications":["If the sample direction set is an $\\varepsilon$-net of the sphere, the Kan-extended PHT is within interleaving distance $2\\varepsilon$ of the true PHT in every homological degree, so denser directional sampling provably improves the approximation.","The extension is 1-Lipschitz, so nearby directions produce nearby persistence modules and the approximated PHT varies continuously over the sphere.","When both directions and height values are sampled, the error is at most $2\\varepsilon_A + \\varepsilon_T$, so the approximation degrades only linearly in the two sampling radii.","The construction works for any constructible shape contained in the unit disk, with no further tameness assumptions beyond o-minimality."],"supporting_citations":[{"why":"Supplies the theorem that coherent maps into persistence modules extend to 1-Lipschitz maps; the interpolation theorem restated as Theorem 3.1.","marker":"[24]"},{"why":"Introduces the space-time category and left/right/center Kan extensions that the paper adapts to the PHT.","marker":"[25]"},{"why":"Introduces the persistent homology transform and proves its Lipschitz stability; the object being approximated.","marker":"[1]"},{"why":"Provides the theory of generalized persistence modules used for the angular $\\Theta$-indexed PHT.","marker":"[26]"},{"why":"Supplies the isometry theorem matching interleaving and bottleneck distances, used in Remark 2.10.","marker":"[29]"},{"why":"Provides the interval-module decomposition theorem underlying the barcode description of persistence modules.","marker":"[28]"}],"fun_headline_variants":["Kan extensions fill in PHT from finite direction samples","PHT interpolation error bounded by direction sample density","Interleaving distance between PHT and extension at most 2ε","Finite directional data recovers PHT with provable error","Sparse directions suffice for persistent homology transform"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approximation only exists if the sample carries the actual homology maps between sublevel sets in different directions—a functor $G$, not just the persistence diagrams—and these maps must be estimated heuristically in practice.","fun_headline_variants_meta":{"raw":{"variants":["Kan extensions fill in PHT from finite direction samples","PHT interpolation error bounded by direction sample density","Interleaving distance between PHT and extension at most 2ε","Finite directional data recovers PHT with provable error","Sparse directions suffice for persistent homology transform"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1453,"prompt_tokens":1069,"completion_tokens":384,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":685,"completion_tokens_details":{"reasoning_tokens":305}},"tokens_in":685,"tokens_out":384,"duration_ms":4907,"temperature":1.0,"reasoning_tokens":305,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:22:12.601589+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the left Kan extension for the single-point shape $M = \\{(0, 1)\\}$ with one sampled direction $v = (1, 0)$, query the south pole $w = (0, -1)$, and compare the birth positions of the true and extended $\\Theta$-persistence modules: the paper's Example 4.5 predicts an interleaving distance of exactly $\\pi$, matching the bound $2 d_g(v, w)$. More generally, a decisive test would evaluate, for a range of constructible shapes inside the unit disk and direction sets $A$, the true interleaving distance between $\\mathrm{PHT}^X_n(M)$ and its center Kan extension and check whether it ever exceeds $2 d_H(A, \\mathbb{S}^{d-1})$ (or $2\\varepsilon_A + \\varepsilon_T$ for fully discrete data).","supporting_citations":[{"cited_title":"& Nanda, V","cited_arxiv_id":null,"evidence_quote":"Supplies the theorem that coherent maps into persistence modules extend to 1-Lipschitz maps; the interpolation theorem restated as Theorem 3.1."},{"cited_title":"& Nanda, V","cited_arxiv_id":null,"evidence_quote":"Introduces the space-time category and left/right/center Kan extensions that the paper adapts to the PHT."},{"cited_title":"Decomposition of pointwise finite-dimension al persistence modules","cited_arxiv_id":null,"evidence_quote":"Provides the interval-module decomposition theorem underlying the barcode description of persistence modules."}],"review_version":1}