{"id":"f42352a0-6285-4894-be86-aa96a4067cc0","arxiv_id":"2507.18170","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A new graphical criterion certifies when causal effects in linear models with arbitrarily connected latent variables are identifiable from the covariance matrix.","lead":"This paper develops a graphical rule that tells when causal effects can be computed from observed data even when hidden variables influence each other in complicated ways. It matters because earlier methods only worked when hidden variables were simple independent factors, which leaves out many realistic scientific models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: Lemma 6.2 and the determinant-nonzero argument in Theorem 3.5 hold up under stress-testing, so the central claim stands.","rationale":"The reader's ACCEPT verdict is reasonable. The central claim is Theorem 3.5, and its proof correctly reduces identifiability to invertibility of the block matrix [A B]. That invertibility is certified by Theorem 6.4, whose key combinatorial input is Lemma 6.2. I examined Lemma 6.2 for hidden exceptions: the monomial equality forces identical edge multisets, and the no-intersection, no-cycle structure of P forces any walk with those edges to follow exactly the same vertex-disjoint paths. The proof's sequential edge-matching argument is complete. The application of Theorem 6.4 in Claim 5 is also faithful: the rows and columns match the definitions, and the required subgraph restrictions are exactly the LSC Condition (iii). The other steps of Theorem 3.5 use standard trek-separation rank bounds and algebraic rank arguments, all of which are consistent with the stated setup. Since no concrete flaw was found in the most load-bearing step, the verdict should remain unchanged. The only residual risk is the lack of a machine-verified proof, which does not by itself warrant a lowered verdict.","tokens_in":27716,"tokens_out":44642,"duration_ms":480791,"concrete_test":"Run an exhaustive search over all directed graphs on up to 6 nodes: enumerate every vertex-disjoint acyclic path system P and every path system Ψ with the same edge multiset; verify that P=Ψ holds in every case, which would settle the determinant-nonzero step behind Theorem 3.5.","verdict_should_be":"UNCHANGED","load_bearing_attack":"No significant objection identified. The reader's weakest assumption is Lemma 6.2, which underlies Claim 5 of Theorem 3.5 via Theorem 6.4. I stress-tested this lemma: if P is a vertex-disjoint acyclic directed path system, equality of monomials forces Ψ to use exactly the same edge set; because P's edge set is a disjoint union of acyclic directed paths with no shared nodes, no walk using those edges can switch paths, form a cycle, or stop short of the prescribed sink without producing a different edge multiset. The induction in the proof is sound. I also checked that Claim 5's block matrix matches Theorem 6.4 with A=Y1, B=Y2, C=pa(v), D=Z, G1=G2=Glat, and that the linear system in Claim 4 has the claimed form. The cited Sullivant et al. trek-separation facts are standard background; no internal inconsistency was found. The remaining uncertainty is only the absence of a machine-checked proof, not a suspected flaw.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a graphical criterion, the latent-subgraph criterion (LSC), for rational identifiability of semi-direct causal effects in linear structural equation models with arbitrarily structured latent variables. The main result, Theorem 3.5, states that if a tuple (Y,Z,H1,H2) satisfies the LSC with respect to an observed node v and all semi-direct effects into a certain set Z ∪ (Y ∩ elr_{H2,H1}(Z ∪ {v})) are already rationally identifiable, then all semi-direct effects p⇝v with p ∈ pa(v) are rationally identifiable. The proof uses the factorization Σ = (I−Λ)^{-⊤}Ω(I−Λ)^{-1}, a new trek-separation-in-subgraphs determinant criterion (Theorem 6.4), and a constructive block linear system. The paper also gives a sound and complete algorithm for checking LSC-identifiability via an integer linear program (Algorithm 1, Theorem 4.7), presents numerical experiments, and compares the new criterion with the canonical model obtained by making all latent nodes source nodes.","tokens_in":27854,"tokens_out":20771,"duration_ms":204820,"significance":"If the main theorem holds, this is a substantial advance: it is, to my knowledge, the first identifiability criterion for semi-direct effects that does not require latent variables to be source nodes, and it applies to arbitrarily structured latent subgraphs. The proof is detailed and constructive, and the determinant-nonzero step is supported by a separate theorem with its own proof. The paper is honest about its limitations: Theorem 6.4 is only sufficient, as Example 6.5 shows, and Conjecture 4.3 about polynomial-time solvability of the integer program is left open. I specifically stress-tested Lemma 6.2, the most fragile combinatorial premise, and found the induction sound. The reproducible code for the simulations is a further strength. The paper is a strong fit for the journal and the central claim is, in my assessment, correct.","major_comments":[],"minor_comments":[{"comment":"In the statement and proof of Lemma 4.5, 'a subset YZ ⊆ Z' should read 'a subset YZ ⊆ Y'; as printed, the claim is nonsensical and inconsistent with Condition (iii) of the LSC.","section":"Lemma 4.5"},{"comment":"Claim 3 states X ⊆ V \\ (Z ∪ {v}), but the expressions Ω_{X,Z}, Ω_{X,v}, and Φ_{X,Z∪{v}} only make sense for X ⊆ O; please state X as a subset of the observed nodes to remove ambiguity.","section":"Appendix A, Claim 3"},{"comment":"The notation in Claim 5 overloads Λ, using it for both the full coefficient matrix and the semi-direct effect matrix; this overloading already appears in Section 2.1. Introducing distinct notation, e.g., Λ̄ for the semi-direct effect matrix, would make the dimension checks in the displayed block matrix transparent.","section":"Appendix A, Claim 5"},{"comment":"There are minor typos: 'no system of of directed paths' in Example 6.5 and 'as we we show' in the discussion preceding it; please correct these.","section":"Section 6 and Section 7"},{"comment":"Equation (1.4) appears typeset incorrectly in the manuscript, with the square root sign and the fraction garbled; please repair the formula.","section":"Example 1.2"},{"comment":"The complexity bound 'O2+kL2k' is not typeset properly; it should presumably read O(2^k |L|^{2k}).","section":"Remark 4.8"}],"recommendation":"accept","confidential_remarks":"The paper is a strong fit for math.ST and the central result appears correct. I see no concern about novelty or scope. The only non-machine-checked combinatorial step, Lemma 6.2, was stress-tested and held up. The minor notational and typographical issues can be fixed during production."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: this paper delivers a real extension, not a repackaging. The latent-subgraph criterion lets you certify rational identifiability of semi-direct effects in linear SEMs where latent variables have arbitrary structure——causal relations among latents, observed-to-latent edges, the works. Previous theory mostly forced latents to be source nodes. The main theorem has a full, constructive proof; the ILP-based checker is a nice practical addition; and the Section 5 observation that canonicalization can flip identifiability is genuinely new and worth citing on its own.\n\nThe paper is honest about its limits. It is sufficient, not necessary; Conjecture 4.3 (polynomial-time via LP instead of ILP) is left open; and NP-hardness and the claim that the criterion strictly subsumes the half-trek criterion are both asserted with sketches rather than full proofs. The trek-separation-in-subgraphs tool, Theorem 6.4, is explicitly only sufficient, and Example 6.5 shows the gap. None of these threaten the central claim.\n\nI stress-tested the one spot the reader worried about: Lemma 6.2, which underlies the determinant-nonzero step in Theorem 3.5 via Theorem 6.4. The lemma holds up; the induction is sound, and the block-matrix identification checks out. The absence of machine-checked proofs is the only real remaining uncertainty, and that is a normal state of affairs for a paper like this.\n\nWho is this for? Anyone working on graphical identifiability, latent variable SEMs, or algebraic statistics. The introduction and examples make it accessible despite the technical density. I would bring it to a reading group and would cite it in my own work. A serious editor should send this to peer review, and the referees should focus on the sufficiency claims and the algorithm's complexity, not on whether the main theorem is true.","headline":"A genuinely new sufficient criterion for identifiability with arbitrary latent structure, honestly limited and largely sound; worth a serious round of review.","tokens_in":28446,"tokens_out":1226,"would_cite":true,"duration_ms":14568,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H05","05C90"],"pacs":[],"model":"deepseek-v4-flash","headline":"A purely graphical criterion certifies when causal effects remain identifiable despite arbitrary latent structure.","keywords":["linear structural equation models","latent variables","causal effect identifiability","semi-direct effects","trek separation","latent-subgraph criterion","rational identifiability","integer linear program"],"falsifier":"Compute the symbolic determinant of the block matrix (A B) from Claim 5 for the graph in Figure 2 (b) with generic symbolic edge weights and the LSC tuple given in Example 3.7; if the determinant simplifies to the zero polynomial, Theorem 3.5 is false for that graph. More directly, search over small directed graphs for two distinct systems of directed paths from the same sources to the same sinks whose monomials are identical, one system having an intersection or a cycle; Lemma 6.2 asserts no such pair exists.","tokens_in":27442,"feed_emoji":"🔗","tokens_out":8901,"duration_ms":81597,"temperature":0.7,"pith_summary":"This paper develops a purely graphical criterion, the latent-subgraph criterion, that certifies when causal effects in a linear structural equation model can be recovered from the observed covariance matrix by explicit rational formulas. The setting is broad: latent variables may influence one another and may mediate effects between observed variables, whereas previous identification criteria typically required every latent variable to be an independent source node. The paper proves the criterion by factoring the covariance matrix around the matrix of semi-direct effects and by introducing a new graph-theoretic tool, trek separation in subgraphs, that certifies the nonvanishing of the determinant blocks needed to solve for the effects. It also supplies an integer-linear-programming algorithm that decides whether the criterion applies, and shows that the standard reduction to a canonical measurement model can change identifiability in both directions.","feed_headline":"Graph criterion certifies causal effects despite hidden causal links","feed_subtitle":"Recovers semi-direct effects by rational formulas, with no assumption that latent variables are independent source nodes.","key_machinery":"The load-bearing object is the latent-subgraph criterion (LSC): a quadruple (Y,Z,H1,H2) satisfying size equalities relating Y to pa(v) and Z to H1,H2; trek separation of Y from Z ∪ {v} inside the latent subgraph Glat (the subgraph of all edges that are not tailed by an observed node); and the existence of a system of treks with no sided intersection from Y to pa(v) ∪ Z whose left parts, and whose right parts ending in Z, use only edges of Glat. The proof of the main theorem rests on a second new object, trek separation in subgraphs (Theorem 6.4), which generalizes the classical trek-separation criterion to block matrices whose entries come from different subgraphs; this is what guarantees that the block matrix being inverted has nonzero generic determinant. The decision algorithm is built from an integer linear program that extends maximum flow to settings where some flows are restricted to a subgraph.","core_discovery":"The central claim is Theorem 3.5: if a 4-tuple (Y,Z,H1,H2) of observed and latent node sets satisfies the latent-subgraph criterion with respect to an observed node v, and all semi-direct effects into the auxiliary nodes Z ∪ (Y ∩ elr_{H2,H1}(Z ∪ {v})) are already known to be rationally identifiable, then all semi-direct effects into v from its semi-direct parents pa(v) are rationally identifiable. Because the criterion is purely graphical, it can be checked column by column: once every observed column is certified, the full semi-direct effect matrix Λ is identified as a rational function of the covariance matrix. The authors emphasize that this is, to their knowledge, the first identifiability criterion for semi-direct effects that does not assume latent nodes are source nodes, and they show by example that identifiability of a model and of its canonicalization are logically independent.","pith_inferences":["The new trek-separation-in-subgraphs condition is not tied to identifiability: a converse characterization of when such block determinants vanish could yield new polynomial constraints on covariance matrices, which in turn could be used for model equivalence and constraint-based testing, as the paper itself suggests in its discussion.","Because the LSC is sufficient but not necessary, one can test how close it is to necessary by comparing it with dimension-based obstructions: for a random sparse graph that fails the LSC, the dimension of the image of the parametrization will often exceed the dimension of the identifiable parameter space, and a systematic comparison would quantify the gap.","The independence of identifiability from canonicalization found in Section 5 suggests that empirical studies that reported non-identifiability of latent-variable structural equation models after canonicalization may have been too pessimistic; rechecking such examples with the LSC may turn some 'unidentified' effects into identified ones."],"forward_implications":["When the criterion certifies every observed column, the full semi-direct effect matrix $\\Lambda$ is rationally identifiable from the observed covariance matrix $\\Sigma$, so each such effect has an explicit closed-form estimator.","Once $\\Lambda$ is known, the residual matrix $\\Omega = (I-\\Lambda)^{\\top}\\Sigma(I-\\Lambda)$ is also identifiable, and $\\Omega$ is the covariance matrix of a simpler measurement model in which effects among latent variables can be identified by existing rules.","The decision problem 'is G LSC-identifiable?' is handled by a sound and complete algorithm that solves integer linear programs; bounding $|H_1|+|H_2|$ makes the number of ILP calls polynomial in $|O|$ and $|L|$, provided Conjecture 4.3 holds, and without the bound the problem is NP-hard.","Confounding-free acyclic graphs are rationally identifiable (Corollary 3.9), giving a broad structural class where the new criterion applies automatically."],"supporting_citations":[{"why":"Supplies the classical trek separation theorem and the determinant expansion used to prove nonvanishing of covariance submatrices; the new subgraph-separation criterion (Theorem 6.4) generalizes it.","marker":"Sullivant et al. (2010)"},{"why":"The latent-factor half-trek criterion that the latent-subgraph criterion subsumes, and whose NP-hardness establishes hardness of the unrestricted decision problem.","marker":"Barber et al. (2022)"},{"why":"The half-trek criterion for linear SEMs whose linear-system derivation underlies the equation system used in Theorem 3.5.","marker":"Foygel et al. (2012)"},{"why":"Defines the canonicalization of a graph; Section 5 compares the original model with this canonical model.","marker":"Hoyer et al. (2008)"},{"why":"Provides the maximum flow problem that the integer linear program in Section 4.1 extends to flows restricted to a subgraph.","marker":"Cormen et al. (2009)"},{"why":"Gives the measurement-model identification rules used to identify effects between latent variables once Λ is recovered.","marker":"Bollen (1989)"}],"fun_headline_variants":["First identifiability criterion for causal effects without source-node latents","Graphical criterion certifies causal identifiability with arbitrary latent structure","Integer programming checks graph criterion for causal effect identification","Certify causal effect identifiability with arbitrarily structured hidden variables"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The identification formula divides by a determinant, and the proof that this determinant is generically nonzero rests on a combinatorial lemma asserting that an intersection-free acyclic system of directed paths contributes a monomial to the determinant that no other path system can reproduce; if that uniqueness lemma had any counterexample, the whole column-by-column identification procedure would fail.","fun_headline_variants_meta":{"raw":{"variants":["First identifiability criterion for causal effects without source-node latents","Graphical criterion certifies causal identifiability with arbitrary latent structure","Integer programming checks graph criterion for causal effect identification","Certify causal effect identifiability with arbitrarily structured hidden variables"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00101,"raw_usage":{"total_tokens":4238,"prompt_tokens":886,"completion_tokens":3352,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":3282}},"tokens_in":502,"tokens_out":3352,"duration_ms":25330,"temperature":1.0,"reasoning_tokens":3282,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:39:10.877924+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the symbolic determinant of the block matrix (A B) from Claim 5 for the graph in Figure 2 (b) with generic symbolic edge weights and the LSC tuple given in Example 3.7; if the determinant simplifies to the zero polynomial, Theorem 3.5 is false for that graph. More directly, search over small directed graphs for two distinct systems of directed paths from the same sources to the same sinks whose monomials are identical, one system having an intersection or a cycle; Lemma 6.2 asserts no such pair exists.","supporting_citations":[{"cited_title":"F., Drton, M., Sturma, N., and Weihs, L","cited_arxiv_id":null,"evidence_quote":"The latent-factor half-trek criterion that the latent-subgraph criterion subsumes, and whose NP-hardness establishes hardness of the unrestricted decision problem."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The half-trek criterion for linear SEMs whose linear-system derivation underlies the equation system used in Theorem 3.5."},{"cited_title":"O., Shimizu, S., Kerminen, A","cited_arxiv_id":null,"evidence_quote":"Defines the canonicalization of a graph; Section 5 compares the original model with this canonical model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the measurement-model identification rules used to identify effects between latent variables once Λ is recovered."}],"review_version":1}