{"id":"7a0a0837-f3dd-4925-ac97-fcb3d7993ac5","arxiv_id":"2412.00847","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A measure-theoretic characterization: X and Y are independent of each other given Z in all product distributions exactly when their histories are almost surely disjoint.","lead":"This paper defines structural independence, the conditional independences that hold for every distribution in which a base family of random elements stays independent. It proves a graph-free criterion using a new object, the history, and sketches an application to causal discovery in structural causal models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 6.2.1's interpolation polynomial is zero at λ=0 and does not connect (P,Q) to (P′,Q′); completeness of Theorem 6.4.1 is therefore unsupported as written.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing flaw: the interpolation polynomial in Theorem 6.2.1 is algebraically wrong, and the proof of completeness depends on that theorem. I agree that this is the most serious obstacle to the central claim. I do not see a reason to tighten the verdict to REJECT: the theorem being proved is plausible, the error appears to be a repairable typo in the interpolation formula rather than a demonstrated counterexample to the classification, and other parts of the paper, such as soundness and the finite theory, do not collapse with this lemma. Conversely, I would not weaken the verdict to ACCEPT: as written, Theorem 6.2.1 is false at the stated polynomial, and the completeness direction of Theorem 6.4.1 is unsupported until the interpolation is corrected and re-verified. The causal application in Section 9 is also informal, but it is downstream of the fundamental theorem and not the primary load-bearing concern. Therefore the reader's CONDITIONAL verdict remains appropriate.","tokens_in":45864,"tokens_out":7094,"duration_ms":66490,"concrete_test":"Re-derive the finite-dimensional step of Theorem 6.2.1 with the corrected interpolation p(λ,x)=∏_{n=0}^m((1−λ)+λx_n) and check three conditions: (i) R_λ and R′_λ are probability measures with R_0=P, R′_0=Q, R_1=P′, R′_1=Q′; (ii) the conditional difference polynomial q(λ,ω)=R_λ(A|Z)R′_λ(B|Z)−R_λ(A,B|Z) satisfies q(0,ω)≠0 on C; (iii) the density dR′_λ/dR_λ is σ(U_i)-measurable for every λ. If any condition fails, Theorem 6.2.1 remains unproven; if all pass, the completeness proof may be salvageable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing defect is the density interpolation in Theorem 6.2.1, used in Theorem 6.2.2 to prove the irrelevance-union statement and then Theorem 6.3.4/6.3.5, i.e. the completeness direction of Theorem 6.4.1. In the finite-dimensional step the paper defines p(λ,x)=∏_{n=0}^m(λx_n+λ) and sets φ_λ=p(λ,Φ), φ′_λ=p(λ,Φ′). But p(0,x)=0 identically, so R_0=φ_0·P is the zero measure rather than P, and R′_0 is not Q. Moreover p(λ,x)=λ^{m+1}∏(x_n+1), so ∫φ_λ dP=(2λ)^{m+1}, which is not 1 for the claimed interpolation; the curve neither consists of probability measures nor connects the required endpoints. The proof then asserts that p′(0,ω)≠0 on C because (R_0,R′_0)=(P,Q)∈△ײ_{i,C}, but p′(0,ω)=0 identically since p(0,·)=0. Thus Theorem 6.2.1 is unproven as stated, and the approximating sequence (P_n,Q_n) used in Theorem 6.2.2 is not delivered. The central characterization may still be true and repairable, e.g. by a convex interpolation such as p(λ,x)=∏((1−λ)+λx_n), but the endpoint, normalization, and σ(U_i)-measurability conditions must be re-established before completeness can be accepted.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a theory of structural independence for an independent family U=(U_i)_{i∈I}. For σ(U)-measurable random elements X, Y, Z, it defines the history H(X|Z), a random index set measuring dependence on U given Z, and claims Theorem 6.4.1: X and Y are independent given Z in every distribution P that renders U independent and is mutually absolutely continuous with respect to the reference measure if and only if H(X|Z)∩H(Y|Z)=∅ almost surely. Sections 4–5 build the machinery of random index sets, disintegration, and generation; Section 6 introduces the dual notion of irrelevance and attempts to prove completeness; Section 7 derives semigraphoid properties and a maximality/uniqueness theorem for the history; Sections 8–9 present a counterexample about disintegration and an application to structural causal models.","tokens_in":46261,"tokens_out":9230,"duration_ms":84707,"significance":"If the completeness proof can be repaired, this is a valuable contribution: it gives a clean random-index-set calculus for distribution-free conditional independence, a simple soundness direction, a uniqueness/maximality theorem, and a semigraphoid structure. The use of Kakutani's theorem in Appendix A is appropriate, and the paper is largely self-contained. The central characterization is attractive. However, the completeness direction is currently unsupported because the key approximation argument in Section 6.2 is invalid as written; this blocks the main theorem.","major_comments":[{"comment":"The interpolation polynomial p(λ,x)=∏_{n=0}^m(λ x_n+λ) does not define the claimed curve. Since p(0,x)=0, one has R_0=0·P, not P, and R'_0=0·P, not Q; moreover ∫φ_λ dP=(2λ)^{m+1}, so R_λ is not a probability measure for λ≠1/2. Consequently the assertion p'(0,ω)≠0 on C is impossible: with this p, p'(0,ω)=0 identically. The approximating sequence (P_n,Q_n) used in Theorem 6.2.2 is therefore not constructed, and the completeness direction of Theorems 6.3.5 and 6.4.1 is unproven as written. A convex interpolation such as p(λ,x)=∏_{n=0}^m((1−λ)+λ x_n) would repair the endpoint and normalization, but this needs to be carried out explicitly and the subsequent polynomial argument re-examined.","section":"§6.2, Theorem 6.2.1"},{"comment":"The equivalence between R_λ(A|Z)=R'_λ(A|Z) and p'(λ,ω)=0 uses the identity E(p(λ,Φ)1_A|Z)=p(λ,(E(φ_n1_A|Z))_n), which is not an instance of linearity of conditional expectation. It would require the factors φ_n to be conditionally independent given Z, which is not assumed and is false in general for σ(U)-measurable Z. As written, the polynomial method does not locate the set where the conditional probabilities differ. The proof of Theorem 6.4.3 in §6.4 appears to rely on the same problematic polynomial reduction.","section":"§6.2, finite-dimensional step"}],"minor_comments":[{"comment":"The statement says that ⋂ J_s 'generates Z given X', and the proof contains the inclusion σ(U_{J_s},Z) ⊆ σ(U_J,Z), which is opposite to the inclusion one would expect from J⊆J_s; this needs correction or clarification.","section":"§5, Lemma 5.9"},{"comment":"In the final sentence, 'Q(A|Z)' should read 'Q(B|Z)'.","section":"§6.1, Theorem 6.1.7"},{"comment":"The symbol p' is used both for the polynomial difference and for a derivative in the phrase p'(0,ω); this is confusing and should be renamed.","section":"§6.2"},{"comment":"The displayed equivalence for A⫫_{R_λ} B|Z has unbalanced parentheses and stray plus signs; as printed it is not correct. The intended algebraic identity should be written out cleanly.","section":"§6.4, Theorem 6.4.3"},{"comment":"The step 'i=ℋ(U_i)⊆ℋ(V_i)' conflates an index with a random index set; the proof of the ancestor conclusion is only sketched and should be made precise.","section":"§9, Lemma 9.4"}],"recommendation":"major_revision","confidential_remarks":"The completeness gap is substantial, not a local typo: the core approximation argument in §6.2 is invalid. I am not recommending rejection because the characterization itself is plausible and the soundness direction is solid, and the errors appear fixable within the manuscript's scope. Editors may also wish to check novelty relative to the author's prior finite theory in [5], [6], and [8], since the paper states Desiderata 5.1.2 as the target characterization before proving it; the new infinite/index-set machinery is the main contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the idea is worth taking seriously: it extends the finite factored-space characterization of structural independence to arbitrary index sets, introduces random index sets (histories and irrelevance), and gives a useful counterexample where rectangular atoms fail. Second, the completeness proof of the main theorem (Theorem 6.4.1) is broken, and the break is not a minor gap. Theorem 6.2.1, the interpolation lemma, defines p(λ,x)=∏_{n=0}^m(λx_n+λ). For λ=0 this is identically zero, so R_0 is the zero measure rather than P. The normalization claim ∫φ_λ dP = 1 is also false; the integral is (2λ)^{m+1} unless the densities are scaled. And p'(0,ω) is zero for m≥1, contradicting the claim that p'(0,ω)≠0 on C. Since Theorem 6.2.1 feeds directly into Theorem 6.2.2 and from there into completeness (6.3.4/6.3.5), the right-to-left direction of the fundamental theorem is unsupported as written.\n\nWhat the paper does well: the setup with random index sets and almost-sure relations is careful and original, and the soundness direction (Theorem 6.1) is straightforward and correct. The counterexample in Section 8 is a useful contribution in its own right. The paper is also honest about what it does not prove, and it does not pretend the infinite case is trivial.\n\nThe soft spots, in proportion: the interpolation error is load-bearing, but it is a technical lemma, not a conceptual one. A convex interpolation like p(λ,x)=∏((1−λ)+λx_n) might repair it, but the endpoint, normalization, and measurability conditions would need to be rechecked, and the approximating sequence in Theorem 6.2.2 would need to be rebuilt. There is also a hand-wavy step in the causal application: Lemma 9.4 says \"it is easy to see\" that history inclusion implies ancestry, but that is exactly the kind of claim that needs proof in an SCM setting. The rest of Section 9 is a sketch.\n\nWho should read this: people working on factored spaces, structural causal models, or distribution-free conditional independence. The paper deserves a serious referee despite the flaw, because the framework is novel and the finite case has already proved useful. My recommendation: send it to review, but with a clear request to fix Theorem 6.2.1 and Lemma 9.4 before the central claim can be accepted. If the interpolation is repaired, the paper would be a solid contribution.","headline":"The infinite-index generalization of structural independence is a real idea, but the completeness proof has a demonstrable error in the interpolation lemma, so the main theorem is unproven as written.","tokens_in":46716,"tokens_out":4563,"would_cite":false,"duration_ms":39042,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60A99"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that an independence is forced by the independence of background variables, in every compatible product distribution, exactly when the two variables' histories are almost surely disjoint.","keywords":["structural independence","conditional independence","random index set","history","d-separation","structural causal model","semigraphoid","infinite product measures"],"falsifier":"Check the finite-dimensional curve in Theorem 6.2.1 directly: the displayed polynomial $p(\\lambda,x)=\\prod_{n=0}^{m}(\\lambda x_n+\\lambda)$ vanishes at $\\lambda=0$, and at $\\lambda=1$ the proposed density $\\varphi_\\lambda$ has total mass $2^{m+1}$ rather than $1$, so the displayed construction is not a probability density; finding a corrected curve that is a probability density for every $\\lambda$ and keeps the conditional-expectation inequality alive for a dense set would repair the proof, while proving no such curve exists would show that the completeness direction is unsupported.","tokens_in":45673,"feed_emoji":"🎲","tokens_out":10966,"duration_ms":96733,"temperature":0.7,"pith_summary":"This paper asks which conditional independences are forced by the bare structural assumption that a family $U=(U_i)_{i\\in I}$ of background random elements is independent, regardless of the numerical distribution they receive. It proves that for any $\\sigma(U)$-measurable random elements $X,Y,Z$, the independence $X\\perp_P Y\\mid Z$ holds for every probability $P$ that makes $U$ independent and is mutually absolutely continuous with the reference measure exactly when the histories $\\mathcal{H}(X\\mid Z)$ and $\\mathcal{H}(Y\\mid Z)$ are almost surely disjoint. The history is a random index set: it picks, for each outcome, the minimal set of coordinates of $U$ on which $X$ still depends after conditioning on $Z$. This gives a $d$-separation-like criterion for structural causal models and settles which independences are structural rather than accidental. A reader should care because it turns a quantification over all distributions into a combinatorial check on random sets.","feed_headline":"Disjoint histories determine which independences are structural","feed_subtitle":"Only overlap-free dependence sets force conditional independence in every product distribution.","key_machinery":"The central object is the history $\\mathcal{H}(X\\mid Z)$, a $\\sigma(Z)$-measurable random index set defined as the almost surely minimal $J(\\omega)\\subseteq I$ such that $\\sigma(X)\\subseteq\\sigma(U_J,Z)$ and $U_J\\perp_P U_{J^c}\\mid Z$ for all $P\\in\\Delta^\\times$. Its dual, the irrelevance $\\mathcal{I}(X\\mid Z)$, is the almost surely maximal set of indices whose one-coordinate perturbations cannot change the conditional distribution of $X$ given $Z$. The proofs of completeness and of the density and rarity results run through the equality $\\mathcal{H}=\\mathcal{I}$ and through a polynomial interpolation argument (Theorem 6.2.1) that connects equivalent product measures by a curve of product densities; the counterexample in Section 8 shows why the finite-case rectangle criterion for disintegration cannot be used in general.","core_discovery":"Theorem 6.4.1 establishes the fundamental theorem of structural independence: for an independent family $U=(U_i)_{i\\in I}$ on $(\\Omega,\\mathcal{A},\\mathbb{P})$, arbitrary $\\sigma(U)$-measurable random elements $X,Y,Z$, and $\\Delta^\\times = \\{P : P\\sim\\mathbb{P},\\ U\\text{ independent under }P\\}$, one has $X\\perp_P Y\\mid Z$ for all $P\\in\\Delta^\\times$ if and only if $\\mathcal{H}(X\\mid Z)\\cap\\mathcal{H}(Y\\mid Z)=\\varnothing$ $\\mathbb{P}$-almost surely. Here $\\mathcal{H}(X\\mid Z)$ is the almost surely minimal $\\sigma(Z)$-measurable random index set that generates $X$ from $U$ given $Z$ and disintegrates $Z$, meaning $U_{\\mathcal{H}}$ and $U_{\\mathcal{H}^c}$ are conditionally independent given $Z$. The paper proves both directions: soundness follows directly from generation, while completeness is obtained by introducing the dual random index set of irrelevance $\\mathcal{I}(X\\mid Z)$, proving $\\mathcal{H}=\\mathcal{I}$, and showing that overlapping histories force a dense set of distributions that violate the independence.","pith_inferences":["The Section 8 counterexample implies that in continuous settings structural independence cannot be read off from the atoms of the conditioning variable; a practical implementation would need reference measures and conditional expectations, not support geometry.","If the interpolation lemma is repaired, the history and irrelevance duality suggests a template for non-product constraint classes: define structural dependence relative to any set of distributions closed under single-coordinate perturbations, and the same minimal and maximal random index set argument might carry through.","The author leaves implicit a statistical reading: because non-structural independences are nowhere dense, random perturbations of an observed distribution should destroy them, which could be turned into a finite-sample test that distinguishes structural from accidental independence in causal discovery.","The equality $\\mathcal{H}=\\mathcal{I}$ indicates that minimal generating set and maximal irrelevant set are two faces of the same quantity; this duality may transfer to other settings where one wants to identify the footprint of a set of variables on an independent noise family."],"forward_implications":["Distribution-free conditional independence for variables built from $U$ reduces to checking whether two random index sets are almost surely disjoint, making a quantified-over-all-distributions statement purely combinatorial.","Structural independence is a compositional semigraphoid: it satisfies symmetry, decomposition, weak union, contraction, and composition, so pairwise structural independence of a vector implies joint structural independence.","In a structural causal model, structural independence is a $d$-separation-like criterion for arbitrary functions of exogenous variables; in the toy setting, $V_1\\perp (V_1+V_2)$ in all compatible models implies $V_1$ is an ancestor of $V_2$, and a parent when $V=(V_1,V_2)$.","If $X$ and $Y$ are not structurally independent given $Z$, then the set of distributions in $\\Delta^\\times$ where $X\\perp_P Y\\mid Z$ holds is closed and nowhere dense in the $d_1$ metric, so accidental, non-structural independences are topologically rare.","The history map is the almost surely unique maximal map satisfying the four desiderata, so it provides a canonical invariant for the structure of an independent family rather than a choice-dependent construction."],"supporting_citations":[{"why":"It supplies the equivalence criterion for infinite product measures used in Lemma A.8 to factor the density between any two distributions in $\\Delta^\\times$ into single-coordinate densities.","marker":"[1]"},{"why":"It provides the $d$-separation soundness and completeness theorem and the structural causal model formalism that the paper generalizes.","marker":"[2]"},{"why":"It gives the original graphical characterization of $d$-separation that structural independence extends to arbitrary random elements.","marker":"[3]"},{"why":"It introduces the finite-case theory of histories, generation, and the rectangle condition that the paper lifts to arbitrary index sets.","marker":"[5]"},{"why":"It provides the finite-case polynomial proof and the toy causal-direction example that the paper adapts to the infinite setting.","marker":"[6]"},{"why":"It supplies the semigraphoid axioms used to show that structural independence is a compositional semigraphoid.","marker":"[9]"}],"fun_headline_variants":["Independence is structural when dependence sets don't meet","Overlapping histories break structural independence","Structural independence: only disjoint histories imply it","Causal direction via disjoint dependence histories","No overlap, no crossing: structural independence's rule"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The completeness half of the fundamental theorem depends on the interpolation lemma in Theorem 6.2.1, which claims that any two equivalent product measures can be joined by a continuous curve of product probability densities while preserving a conditional-probability inequality; if this lemma cannot be repaired, the proof that overlapping histories force failure of independence in some distribution is incomplete.","fun_headline_variants_meta":{"raw":{"variants":["Independence is structural when dependence sets don't meet","Overlapping histories break structural independence","Structural independence: only disjoint histories imply it","Causal direction via disjoint dependence histories","No overlap, no crossing: structural independence's rule"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000206,"raw_usage":{"total_tokens":1467,"prompt_tokens":1085,"completion_tokens":382,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":701,"completion_tokens_details":{"reasoning_tokens":315}},"tokens_in":701,"tokens_out":382,"duration_ms":3856,"temperature":1.0,"reasoning_tokens":315,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:56:33.273560+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check the finite-dimensional curve in Theorem 6.2.1 directly: the displayed polynomial $p(\\lambda,x)=\\prod_{n=0}^{m}(\\lambda x_n+\\lambda)$ vanishes at $\\lambda=0$, and at $\\lambda=1$ the proposed density $\\varphi_\\lambda$ has total mass $2^{m+1}$ rather than $1$, so the displayed construction is not a probability density; finding a corrected curve that is a probability density for every $\\lambda$ and keeps the conditional-expectation inequality alive for a dense set would repair the proof, while proving no such curve exists would show that the completeness direction is unsupported.","supporting_citations":[{"cited_title":"On equivalence of infinite product measures,","cited_arxiv_id":null,"evidence_quote":"It supplies the equivalence criterion for infinite product measures used in Lemma A.8 to factor the density between any two distributions in $\\Delta^\\times$ into single-coordinate densities."},{"cited_title":"Causal networks: Semantics and expressiveness,","cited_arxiv_id":null,"evidence_quote":"It gives the original graphical characterization of $d$-separation that structural independence extends to arbitrary random elements."},{"cited_title":"Factored Space Models: Towards Causality between levels of abstraction ,","cited_arxiv_id":null,"evidence_quote":"It introduces the finite-case theory of histories, generation, and the rectangle condition that the paper lifts to arbitrary index sets."},{"cited_title":"Graphoids: Graph-Based Logic for Reasoning about Relevance Relations or When would x tell you more about y if you already know z?,","cited_arxiv_id":null,"evidence_quote":"It supplies the semigraphoid axioms used to show that structural independence is a compositional semigraphoid."}],"review_version":1}