{"id":"e0b83282-cf24-471a-9843-b8c98a50abdb","arxiv_id":"2506.19991","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For two embeddings of the same finite simplicial complex, the integrated distance between their Euler Characteristic Transforms is bounded by a constant times the total vertex displacement, and SELECT obeys an analogous bound for simplex-wise monotone fields.","lead":"This paper proves an upper bound on how much the Euler Characteristic Transform changes when the same abstract shape is embedded in space in two slightly different ways, provided the combinatorial mesh is fixed. It also proves a similar bound for SELECT, a variant that reads scalar fields on shapes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SELECT bound is formally proven only for piecewise-constant fields; the text's retraction bridge to PL fields is false for halfspace-truncated superlevel sets, so the scope claim is too broad.","rationale":"The ECT theorem (Theorem 4.1) is a clean composition of the cited Dlotko-Gurnari and Skraba-Turner bounds, and the formal SELECT theorem for piecewise-constant fields follows from it; I do not see a defect there. The load-bearing problem is the paper's unproved bridge from piecewise-constant to PL fields. The reader's weakest assumption identifies exactly this, and a concrete triangle computation shows the bridge is not just unproved but false for SELECT's height-truncated intersections. Since the Discussion explicitly lists PL interpolation as future work, the correct remedy is a scope clarification and removal or correction of the retraction paragraph, not rejection. This keeps the reader's CONDITIONAL verdict unchanged. Minor issues such as the nonexistent 'Lemma 3.2' cross-reference and the omitted proof of Lemma 3.6 are expositional and do not affect the main bounds.","tokens_in":11399,"tokens_out":25738,"duration_ms":281675,"concrete_test":"Run the one-triangle example with v1=(0,0), v2=(2,0), v3=(2,2), φ(v1)=φ(v2)=1, φ(v3)=2. For ν=(1,0), a=1.5, t=1.5, evaluate both sides of the claimed replacement. The piecewise-constant K_t={v3} gives ECCν(f(K_t),1.5)=0. The PL interpolation Φ=1+λ_{v3} has {x≤1.5, Φ≥1.5} nonempty with χ=1. The two values differ, so the retraction argument cannot justify replacing PL fields by piecewise-constant superlevel sets in SELECT.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 5.3 is stated for fields of the form SELECT(φ∘f^{-1}) with φ simplex-wise monotone, i.e. piecewise-constant superlevel sets, and that proof is internally sound. The problem is the bridge the paper tries to build to PL fields. The paragraph after Definition 5.2 says a PL interpolation Φ of φ has Φ^{-1}[t,∞) homotopy equivalent to K_t, 'and so we are justified in using the piecewise constant function instead.' This is false for the height-truncated intersections SELECT actually evaluates. Concretely, let K be the single triangle with v1=(0,0), v2=(2,0), v3=(2,2), φ(v1)=φ(v2)=1, φ(v3)=2, direction ν=(1,0), a=1.5, t=1.5. Then K_t={v3}, so ECCν(f(K_t),1.5)=0, since hν(v3)=2. But the PL function Φ=1+λ_{v3} has {Φ≥1.5} equal to the subtriangle near v3 (λ_{v3}≥0.5), and this subtriangle meets the halfspace {x≤1.5} in a nonempty set of Euler characteristic 1. Thus SELECT(Φ)(ν,a,t)=1 while the piecewise-constant replacement gives 0. The homotopy equivalence of the full superlevel sets does not survive truncation by {⟨x,ν⟩≤a}. Consequently, the paper has not proved SELECT stability for PL fields, despite Definition 5.2 presenting SELECT for PL functions and the abstract saying 'fields defined on embedded simplicial complexes.' This should be fixed by explicitly scoping Theorem 5.3 to piecewise-constant/simplex-wise monotone fields and deleting or repairing the retraction paragraph; the Discussion already lists PL interpolation as future work, which is the correct honest scope.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies stability of the Euler characteristic transform (ECT) and its field-valued variant SELECT under perturbations of an embedding of a fixed abstract simplicial complex K into R^d. Theorem 4.1 bounds the integrated L1 distance between the ECTs of two embeddings by 2 C_K C_d Σ_v ||f(v)-g(v)||_2, chaining a bound of Dlotko and Gurnari with the Wasserstein stability theorem of Skraba and Turner. For SELECT, the paper assigns vertex values φ, forms superlevel subcomplexes K_t, and proves Theorem 5.3, an analogous bound with an extra factor 2 r_max obtained by integrating Theorem 4.1 over t. The paper also proves two metric lemmas and concludes with a discussion of limitations, including the fact that piecewise-linear interpolation is left to future work.","tokens_in":11829,"tokens_out":9793,"duration_ms":99178,"significance":"The ECT stability result is a clean and useful contribution: it provides an explicit Lipschitz-type bound with constants depending only on the complex and the dimension, and the proof is a transparent chain of published theorems. The SELECT result is also valuable for piecewise-constant, simplex-wise monotone fields, where Theorem 5.3 gives an explicit bound with a natural dependence on the field range. However, the paper's advertised scope is broader than what is proved: the bridge used to pass from PL interpolation to piecewise-constant superlevel sets is invalid for the halfspace-truncated sets that SELECT evaluates, so the SELECT contribution needs to be explicitly scoped before the claims in the abstract and Definition 5.2.","major_comments":[{"comment":"This paragraph asserts that a PL interpolation Φ of a simplex-wise monotone function has Φ^{-1}[t,∞) homotopy equivalent to K_t and therefore 'we are justified in using the piecewise constant function instead.' The equivalence is not enough: SELECT evaluates χ({x: ⟨x,ν⟩≤a, Φ(x)≥t}), and homotopy equivalence of the untruncated superlevel set does not survive intersection with the halfspace {⟨x,ν⟩≤a}. Concretely, take K to be the single triangle with vertices (0,0), (2,0), (2,2), φ(v1)=φ(v2)=1, φ(v3)=2, ν=(1,0), a=1.5, t=1.5. Then K_t={v3}, so the piecewise-constant replacement gives ECCν(f(K_t),1.5)=χ(∅)=0. The PL interpolation is Φ=1+λ_{v3}; its superlevel set {Φ≥1.5} is a subtriangle near v3, and its intersection with {x≤1.5} is nonempty and convex, so SELECT(Φ)(ν,a,t)=1. Hence Theorem 5.3 is not proved for PL fields. The theorem should be explicitly restricted to piecewise-constant/simplex-wise monotone fields, and the retraction paragraph should be repaired or removed; the Discussion already correctly lists PL interpolation as future work.","section":"Section 5, paragraph beginning 'However, since this idea...'"},{"comment":"There is a domain mismatch in the statement of Theorem 5.3: SELECT is defined in Definition 5.2 only for PL functions, but φ∘f^{-1} is piecewise constant and generally discontinuous, so it is not an element of PL(R^d). In addition, f^{-1}(x) is not uniquely defined for points on shared faces of simplices, and the value of φbar can differ on a face and on a higher-dimensional coface. The theorem should define ef and eg directly from the superlevel subcomplexes, for example ef(ν,a,t)=ECCν(f(K_t),a), and describe the fields as piecewise-constant fields associated to φ rather than as SELECT of a PL function.","section":"Theorem 5.3 and Definition 5.2"}],"minor_comments":[{"comment":"The reference to 'Lemma 3.2' should be to Theorem 3.3, which is the inequality W_{1,∞}≤W_{1,1} used in that proof.","section":"Proof of Theorem 4.1"},{"comment":"Since φ is defined on vertices, the quantity r_max should be defined as max_{σ∈K} min_{v∈σ} φ(v) (or as the maximum of the vertex values), rather than max_{σ∈K} φ(σ), to avoid a domain mismatch.","section":"Section 5, definition of r_max"},{"comment":"There are several typos: 'continued' should be 'continuous', 'monoanalogous' should be 'monotone', and the definition 'ef:S^{d-1}×R×R' is missing its codomain.","section":"Section 5, text near Definition 5.2"},{"comment":"After scoping Theorem 5.3, the phrases 'fields defined on embedded simplicial complexes' in the abstract and 'the use of a particular class of field-inducing functions' in the Discussion should be made precise by referring to piecewise-constant, simplex-wise monotone fields.","section":"Abstract and Section 6"}],"recommendation":"major_revision","confidential_remarks":"The only substantive correctness concern is the false bridge from PL interpolation to piecewise-constant superlevel sets in the SELECT section. The paper's own Discussion already lists PL interpolation as future work, so the fix is to align the theorem statement, definitions, and abstract with what is actually proved. I see no issues with novelty or citation practice."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the ECT half is a correct, clean composition of known inequalities, and the explicit Lipschitz constant is worth having. The SELECT half is a real extension but is proven only for piecewise-constant, simplex-wise monotone fields. The paragraph that tries to bridge to PL functions asserts a homotopy equivalence that does not survive the halfspace truncation SELECT evaluates; the stress-test counterexample with a single triangle is correct. So the SELECT claim should be explicitly scoped to piecewise-constant fields, with PL interpolation left as future work—which the Discussion already lists.\n\nWhat is actually new and good: Theorem 4.1 states an explicit bound with constants C_K and C_d, chaining Dlotko-Gurnari, Skraba-Turner, and Turner with all citations present and no hidden parameters. The proof is transparent. The SELECT bound's reduction of d_SELECT to an integral of d_ECT over t is neat, and the steps C_{K_t} ≤ C_K and V(K_t) ⊆ V(K) are correct. The writing is clear, and the figures borrowed from Munch are well chosen.\n\nSoft spots, in order of importance. First, the PL bridge. The paragraph beginning \"However, since this idea is closely related...\" claims that Φ^{-1}[t,∞) retracts to K_t and therefore the piecewise-constant function can be used instead. The homotopy equivalence of the full superlevel sets is true, but SELECT evaluates {x : ⟨x,ν⟩ ≤ a, Φ(x) ≥ t}, and truncation by the halfspace breaks the equivalence. The concrete counterexample in the stress-test note is exactly right: the truncated subtriangle has Euler characteristic 1 while the piecewise-constant replacement gives 0. So Theorem 5.3 is not established for PL fields. This is a scope problem, not a fatal one, because the proof is sound for the piecewise-constant fields it actually uses, and the Discussion already names PL interpolation as future work. The authors should delete or repair that paragraph and state the theorem for simplex-wise monotone fields only.\n\nSecond, the proof of Lemma 3.6 is omitted. Minor; it really is almost identical to Lemma 3.5. Third, the proof of Theorem 4.1 cites \"Lemma 3.2,\" which does not exist; it should cite Theorem 3.3 (Turner's W_{p,∞} ≤ W_{p,p}). Minor typo, but confusing.\n\nCitation pattern: no self-citation issues. Munch [15] is used only as a survey and figure source, which is appropriate. The external theorems are all properly credited.\n\nWho this is for: TDA practitioners who want stability guarantees for deformable objects with fixed triangulation, and anyone building ML pipelines on ECT or SELECT. The bound is not tight but explicit, which is what practice needs.\n\nRecommendation: send it to peer review. It deserves a serious referee. I would accept after minor revision, with the SELECT scope fixed and the two smaller issues cleaned up.","headline":"The ECT Lipschitz bound is a clean, correct composition of existing theorems, but the SELECT bound is proven only for piecewise-constant fields and the paper's attempt to extend it to PL functions rests on a false homotopy-retraction claim.","tokens_in":12355,"tokens_out":2516,"would_cite":true,"duration_ms":25377,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["55N31","68U05"],"pacs":[],"model":"deepseek-v4-flash","headline":"For two embeddings of the same simplicial complex, the Euler Characteristic Transform distance is at most a constant times the total vertex displacement.","keywords":["Euler characteristic transform","SELECT","simplicial complex","stability","Wasserstein distance","persistence diagram","topological data analysis","piecewise linear field"],"falsifier":"For the ECT bound, take the simplest nontrivial complex, a single edge in $\\mathbb{R}^2$, fix embeddings $f$ and $g$ that differ by moving one endpoint a distance $\\varepsilon$, and compute $d_{\\mathrm{ECT}}$ exactly; the theorem fails if it exceeds $2C_KC_d\\varepsilon$. For the SELECT bound, take a genuine piecewise-linear interpolation $\\Phi$ of vertex weights on a triangle, move one vertex by $\\varepsilon$, and compare $d_{\\mathrm{SELECT}}(\\Phi\\circ f^{-1},\\Phi\\circ g^{-1})$ with $2r_{\\max}C_KC_d\\varepsilon$; exceeding that value would show the asserted replacement of PL fields by piecewise-constant fields does not preserve the bound.","tokens_in":11213,"feed_emoji":"📐","tokens_out":11150,"duration_ms":104243,"temperature":0.7,"pith_summary":"In plain terms, the paper proves that two quantitative shape descriptors from topological data analysis are stable under small changes in the geometric embedding of a fixed simplicial complex. For the Euler Characteristic Transform, the result is an explicit bound: the distance between the transforms of two embeddings $f$ and $g$ of the same abstract complex $K$, measured by integrating the $L^1$ difference of their Euler characteristic curves over all viewing directions, is at most $2\\,C_K C_d \\sum_{v\\in V(K)}\\|f(v)-g(v)\\|_2$. For the Super Lifted Euler Characteristic Transform, defined for scalar fields on shapes, a similar bound holds with an extra factor equal to the range of the field, when the field is piecewise constant and simplex-wise monotone. This matters because the ECT is already used in shape classification and machine-learning pipelines, and a proven stability bound says that small vertex perturbations cannot arbitrarily distort the descriptor.","feed_headline":"Moving vertices changes the shape transform by at most a fixed factor","feed_subtitle":"A proven Lipschitz bound makes the transform reliable under mesh deformations.","key_machinery":"The load-bearing machinery is a chain of Wasserstein comparisons. The paper builds the ECT distance as $\\int_{S^{d-1}}\\|\\mathrm{ECC}_\\nu(f(K))-\\mathrm{ECC}_\\nu(g(K))\\|_1\\,d\\nu$, then bounds the integrand by $2W_{1,\\infty}$ of the persistence diagrams of the two height filtrations, uses a known direction-integrated $W_{1,1}$ stability bound whose constants are $C_K$ (the maximum number of simplices incident to a vertex) and $C_d$ (a fixed sphere integral), and closes the chain with the inequality $W_{1,\\infty}\\le W_{1,1}$. For SELECT, the mechanism is that superlevel sets $K_t=\\{x:\\phi(x)\\ge t\\}$ of a simplex-wise monotone field are subcomplexes, so $\\mathrm{SELECT}(\\phi\\circ f^{-1})(\\nu,a,t)=\\mathrm{ECC}_\\nu(f(K_t),a)$, turning each fixed-$t$ slice into an ECT computation and leaving a single integral over $t\\in[0,r_{\\max}]$.","core_discovery":"On its own terms, the central discovery is that the ECT and SELECT are Lipschitz functions of the vertex positions of a triangulation. Theorem 4.1 states that for two embeddings $f,g$ of the same abstract simplicial complex $K$ into $\\mathbb{R}^d$, the distance $d_{\\mathrm{ECT}}(\\mathrm{ECT}(f(K)),\\mathrm{ECT}(g(K)))$, defined by integrating the $L^1$ distance of Euler characteristic curves over all directions $\\nu\\in S^{d-1}$, satisfies $d_{\\mathrm{ECT}}\\le 2C_KC_d\\sum_{v\\in V(K)}\\|f(v)-g(v)\\|_2$, where $C_K$ counts the maximum number of simplices containing any vertex and $C_d$ is a constant depending only on $d$. The proof chains three existing inequalities: an Euler-characteristic-curve difference is bounded by a Wasserstein distance between persistence diagrams, that Wasserstein distance integrated over all directions is bounded by a constant times total vertex displacement, and the two Wasserstein metrics are comparable. Theorem 5.3 gives an analogous bound for SELECT: for a simplex-wise monotone vertex function $\\phi$, the SELECT distance between $\\mathrm{SELECT}(\\phi\\circ f^{-1})$ and $\\mathrm{SELECT}(\\phi\\circ g^{-1})$ is at most $2r_{\\max}C_KC_d\\sum_{v}\\|f(v)-g(v)\\|_2$, by recognizing that SELECT at threshold $t$ is exactly the ECT of the superlevel subcomplex $K_t$ and integrating the ECT bound over $t$ up to $r_{\\max}$.","pith_inferences":["The same chain of inequalities would likely prove stability for any transform that controls its per-direction output by a Wasserstein distance between persistence diagrams; the authors do not state this general template.","The ECT bound implies a statistical continuity the paper leaves implicit: if vertex positions are estimated with small registration error, empirical ECT-based summaries change by at most a fixed multiple of that error, so confidence statements about shape descriptors inherit continuity in the mesh.","The factor $r_{\\max}$ in the SELECT bound is plausibly loose, since $K_t$ shrinks as $t$ grows and $C_{K_t}$ decreases; a sharper bound integrating only over the actual spread of vertex values might be strictly smaller.","A closed-form check on a single edge or triangle would calibrate how much slack the constants $C_K$ and $C_d$ carry, since the paper gives no optimality example."],"forward_implications":["If Theorem 4.1 is correct, the ECT is a Lipschitz map from the space of embeddings of a fixed abstract complex (metrized by total vertex displacement) to the space of transforms with the direction-integrated $L^1$ distance.","Small geometric perturbation of a shape whose triangulation is unchanged changes the ECT by at most a constant times the sum of vertex movements, with the constant computable from the complex and the dimension.","The SELECT bound extends the same guarantee to scalar fields: for step-like fields satisfying the simplex-wise monotone condition, moving the embedding changes SELECT by at most a constant times the total vertex displacement times the field range $r_{\\max}$.","Because all constants are explicit, a practitioner can predict the worst-case descriptor drift from a known mesh deformation before recomputing the transform."],"supporting_citations":[{"why":"Supplies the inequality that bounds the $L^1$ distance of Euler characteristic curves by the Wasserstein distance of the associated persistence diagrams, the first link in the ECT proof.","marker":"[6]"},{"why":"Gives the integrated Wasserstein stability bound for persistence diagrams of height filtrations on two embeddings of the same complex, the central input to Theorem 4.1.","marker":"[19]"},{"why":"Provides the comparison $W_{1,\\infty}\\le W_{1,1}$ between Wasserstein distances that lets the first two results be chained.","marker":"[21]"},{"why":"Defines the ECT as a transform and supplies the metric structure and injectivity background on which the ECT distance is built.","marker":"[4]"},{"why":"Defines LECT and SELECT and proves their injectivity, giving the object whose stability is the subject of Theorem 5.3.","marker":"[11]"}],"fun_headline_variants":["ECT Lipschitz in vertex position with explicit constant","Vertex moves cause at most linear change in ECT","ECT and SELECT stability: bound by sum of vertex displacements","Shape transform changes linearly with vertex perturbation","Proven Lipschitz bound for Euler Characteristic Transform"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The SELECT result is proved only for step-like fields that are constant on each simplex; the paper's argument that this also covers ordinary piecewise-linear interpolation depends on an asserted homotopy equivalence between the PL superlevel sets and the simplicial superlevel sets, and the paper lists the PL case itself as future work.","fun_headline_variants_meta":{"raw":{"variants":["ECT Lipschitz in vertex position with explicit constant","Vertex moves cause at most linear change in ECT","ECT and SELECT stability: bound by sum of vertex displacements","Shape transform changes linearly with vertex perturbation","Proven Lipschitz bound for Euler Characteristic Transform"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000594,"raw_usage":{"total_tokens":2818,"prompt_tokens":1014,"completion_tokens":1804,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":1729}},"tokens_in":630,"tokens_out":1804,"duration_ms":13941,"temperature":1.0,"reasoning_tokens":1729,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:22:46.355524+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the ECT bound, take the simplest nontrivial complex, a single edge in $\\mathbb{R}^2$, fix embeddings $f$ and $g$ that differ by moving one endpoint a distance $\\varepsilon$, and compute $d_{\\mathrm{ECT}}$ exactly; the theorem fails if it exceeds $2C_KC_d\\varepsilon$. For the SELECT bound, take a genuine piecewise-linear interpolation $\\Phi$ of vertex weights on a triangle, move one vertex by $\\varepsilon$, and compare $d_{\\mathrm{SELECT}}(\\Phi\\circ f^{-1},\\Phi\\circ g^{-1})$ with $2r_{\\max}C_KC_d\\varepsilon$; exceeding that value would show the asserted replacement of PL fields by piecewise-constant fields does not preserve the bound.","supporting_citations":[{"cited_title":"Euler characteristic curves and profiles: a stable shape invariant for big data problems.GigaScience, 12:giad094, 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the inequality that bounds the $L^1$ distance of Euler characteristic curves by the Wasserstein distance of the associated persistence diagrams, the first link in the ECT proof."},{"cited_title":"Medians of populations of persistence diagrams.Homology, Homotopy and Applications, 22(1):255–282, July 2020","cited_arxiv_id":null,"evidence_quote":"Provides the comparison $W_{1,\\infty}\\le W_{1,1}$ between Wasserstein distances that lets the first two results be chained."},{"cited_title":"Representing fields without correspondences: the lifted Euler characteristic transform.Journal of Applied and Computational Topology, 8(1):1–34, 2024","cited_arxiv_id":null,"evidence_quote":"Defines LECT and SELECT and proves their injectivity, giving the object whose stability is the subject of Theorem 5.3."}],"review_version":2}