{"id":"007edd93-00c2-49da-9c1d-a0279f5d888f","arxiv_id":"2505.00215","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A random vector arises from a linear non-Gaussian acyclic model with graph G if and only if certain matrices built from its second- and third-order moments have rank exactly equal to the number of parents of each vertex.","lead":"This paper proves a complete algebraic test for whether a random vector's second- and third-order moments can come from a linear non-Gaussian causal model with a given directed acyclic graph. The test is a set of easily checked rank conditions on moment matrices, generalizing known results for trees and for Gaussian models.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified","rationale":"I read Theorem 3.1 as the central claim: for any DAG, membership in M<=3(G) is exactly the rank condition on the matrices M_v. I checked the necessity direction, the construction of Lambda in Lemma 3.6, and the diagonalization of Omega^(2) and Omega^(3). The proof is mathematically coherent once one corrects a clear index typo in the Omega^(2) expansion: the inner sum over u must range over pa(v), not pa(w). The Omega^(3) calculation is valid because the row relation for M_v holds for all third indices, and the symmetry of T lets one place a non-descendant in the second slot. I also checked edge cases such as sources with empty parent sets and pairs where one vertex is a descendant of the other; the use of symmetry covers these. I found no counterexample and no missing algebraic condition. The reader's weakest assumption concerns the LiNGAM independence and no-unobserved-confounding assumptions; this is genuine for applications but it is not a flaw in the theorem, which is explicitly conditional on the model. The numerical experiments are illustrative and lack code and data, but they are not load-bearing for the algebraic characterization. The vanishing ideal conjecture is acknowledged as open and does not affect Theorem 3.1. Overall, the central claim survives scrutiny, so the ACCEPT verdict should remain unchanged.","tokens_in":11500,"tokens_out":28559,"duration_ms":301536,"concrete_test":"Re-run the proof of Lemma 3.6 with the corrected index u in pa(v) and independently verify the Macaulay2 example in Example 3.3 on a four-vertex non-polytree DAG: compute the kernel of the parametrization and compare it with the ideal generated by the (|pa(v)|+1)-minors of the matrices M_v, saturated by principal minors of S. Equality of these ideals confirms the rank characterization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"No significant objection identified. The central rank characterization in Theorem 3.1 is internally sound: necessity follows from the LiNGAM moment equations because a proper non-descendant X_w is independent of the error term, and sufficiency constructs Lambda from the unique row relations in M_v and verifies that the transformed residual covariance and third-order tensor are diagonal. The only caveats are editorial: the displayed expansion of Omega^(2) in Lemma 3.6 uses u in pa(w) where the correct index is u in pa(v), and the zero-third-moment argument for T_{v,w,z} silently uses E[X_w]=0 when z is a descendant of v. Both are repaired without changing the theorem. The reader's caveat about error independence is a modeling assumption, not a defect of the theorem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies the set of pairs (S,T) of covariance matrices and third-order moment tensors arising from linear non-Gaussian acyclic models (LiNGAM) with a given DAG G. The main result (Theorem 3.1) states that (S,T) belongs to the model M≤3(G) if and only if for every vertex v, the matrix M_v defined in equation (2), built from S and T blocks involving parents and non-descendants of v, has rank exactly |pa(v)|. The necessity argument uses the independence of the error at v from all non-descendant variables; the sufficiency argument constructs the coefficient matrix Λ from the row dependencies of M_v and verifies that the transformed residual second- and third-order moments are diagonal. The paper also derives additional equations using a generalized parent/non-descendant operation (Proposition 4.2), recovers prior polytree results, and presents an application to sink detection.","tokens_in":11626,"tokens_out":25958,"duration_ms":226841,"significance":"Theorem 3.1, if correct, provides the first complete algebraic characterization of the second- and third-order moment model of LiNGAM for an arbitrary DAG, extending the polytree case of ADG+23 and containing the Gaussian local Markov rank constraints. The proof is self-contained, uses only standard linear algebra, and the constraints are explicit determinantal conditions that can be directly computed from moments. The paper includes symbolic verification in Macaulay2 for small graphs and reproducible simulations. However, the application section contains a false mathematical claim about sink recovery, overstating what Theorem 3.1 implies for structure learning.","major_comments":[{"comment":"The sentence 'By Theorem 3.1, the true matrix M_i will drop rank if and only if i is not an ancestor of any other vertices in the graph' is false. Counterexample: let G be the complete DAG on {1,2,3} with edges 1→2, 1→3, 2→3, and let X1=ε1, X2=ε1+ε2, X3=ε1+ε2+ε3, where ε_i are i.i.d. with mean 0, variance 1 and third moment 1. For i=2, which is an ancestor of 3, the matrix in (11) has rows indexed by 1,3,2; direct computation gives row2 = 0.5 row1 + 0.5 row3 for every column, so its rank is 2 < 3. Thus a non-sink can produce a rank-deficient matrix, contradicting the claimed 'only if' direction. This does not affect the validity of Theorem 3.1, but it means the sink-recovery algorithm is not justified as stated; the authors should replace the claim with the correct one-directional statement (a sink implies rank deficiency) and present the method as a heuristic, or prove a different characterization.","section":"Section 6, equation (11)"}],"minor_comments":[{"comment":"In the displayed expansions of Ω^(2) and Ω^(3), the inner summation index is written as u∈pa(w) but should be u∈pa(v); otherwise the equations are inconsistent with the row relations.","section":"Section 5, proof of Lemma 3.6"},{"comment":"The step 'A similar equation can be derived for t_{vwz}' implicitly uses E[ε_v X_w X_z]=0, which holds because E[X_w]=0 and ε_v is independent of X_w; this justification should be added.","section":"Lemma 3.5"},{"comment":"The text says 'without loss of generality' errors have mean zero, but the simulations in Section 6 sample errors from Γ(5,1), which has mean 5; if the errors are not centered, E[X_i]≠0. Please clarify that the sampled errors are centered before use.","section":"Section 2.1 and Section 6"},{"comment":"There are several typos: 'knwon' in Section 2.2, 'poof' in Remark 5.1, and 'Learninig' in reference [SHHK06].","section":"Typos"},{"comment":"The caption 'The graph is a bar chart, each bar represents 200 data points' is unclear; please describe how the bar chart encodes sample sizes and proportions.","section":"Figure 3 caption"}],"recommendation":"minor_revision","confidential_remarks":"The central theorem is sound and the paper is a strong contribution. The main issue is the false claim in Section 6, which should be corrected; it does not affect the main result. I recommend minor revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Have you seen arXiv:2505.00215? The headline: the main theorem is correct and it is a genuine completion of an open characterization. For any DAG G, M≤3(G) is cut out by rank conditions on the matrices M_v built from S and T. Previous work had polytrees (ADG+23) and Gaussian covariance constraints; this paper covers all DAGs and shows the Gaussian local Markov constraints are a sub-block. That is a real result, not a renaming of earlier equations.\n\nThe proof structure is right. Necessity (Lemma 3.5) uses independence between ε_v and proper non-descendants to get a left null vector; sufficiency (Lemma 3.6) reconstructs Λ from the unique row combination and verifies that the residual covariance and third-order tensor are diagonal. I checked the two small issues the stress-test flagged: the expansion in Lemma 3.6 writes u∈pa(w) where pa(v) is meant, and the zero-third-moment argument silently uses E[X_w]=0 when z is a descendant of v. Both are cosmetic; the repair is immediate and doesn't touch the theorem. Section 4's derivation of polytree equations from Proposition 4.2 is a nice unification, and Corollary 4.4 explains the WD19 source-recovery check as a special case.\n\nSoft spots, in proportion:\n\n- The numerical experiments in Section 6 are illustrative. No code or data is shipped, the graphs are tiny, and the ROC analysis is a threshold sweep on simulated data. That is fine for a theory paper, but claims about practical sink recovery should be read as suggestive, not evaluated.\n- The vanishing ideal is left as a conjecture, and the paper says so explicitly. Not a flaw; just a boundary.\n- The theorem depends on the LiNGAM error assumptions. If errors are dependent or there is unobserved confounding, the moment equations break and the rank condition loses its graph-theoretic reading. The paper is upfront about the model, and this limitation is inherent to the setting, not a gap in the proof.\n\nCitation pattern looks normal: prior work is credited, and self-citations are to the relevant polytree and moment-variety papers. No red flag.\n\nI would send this to a serious referee. It proves what it claims and closes a question that was explicitly open. My recommendation: accept after minor revision, with the typos in Lemma 3.6 fixed and the experiments labeled as illustrative. I would cite it.","headline":"The main theorem is correct and closes an open characterization: for any DAG, the LiNGAM second- and third-order moment model is cut out by explicit rank conditions on the matrices M_v.","tokens_in":12118,"tokens_out":1848,"would_cite":true,"duration_ms":18145,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62R01"],"pacs":[],"model":"deepseek-v4-flash","headline":"A random vector comes from a LiNGAM on a DAG exactly when certain moment matrices have rank equal to the number of parents of each vertex.","keywords":["LiNGAM","causal discovery","third-order moments","algebraic statistics","determinantal constraints","structural equation models","directed acyclic graph","method of moments"],"falsifier":"The theorem would be refuted by a triple $(G,S,T)$ with $S$ positive definite and $T$ symmetric where every matrix $M_v$ has rank $|\\mathrm{pa}(v)|$ but $(S,T)$ is not in the image of the parametrization $S=(I-\\Lambda)^{-T}\\Omega^{(2)}(I-\\Lambda)^{-1}$, $T=\\Omega^{(3)}\\bullet (I-\\Lambda)^{-1}\\bullet (I-\\Lambda)^{-1}\\bullet (I-\\Lambda)^{-1}$ for any $\\Lambda\\in\\mathbb{R}^E$ and diagonal $\\Omega^{(2)},\\Omega^{(3)}$. For the three-vertex DAG $1\\to 2$, $1\\to 3$, $2\\to 3$, one can test this directly with a computer algebra system: the theorem predicts the model is cut out by the 2-minors of $M_2$ and 3-minors of $M_3$ saturated by the principal minors of $S$, so any point in that variety outside the parametrization image would disprove the sufficiency direction.","tokens_in":11317,"feed_emoji":"📐","tokens_out":8629,"duration_ms":81558,"temperature":0.7,"pith_summary":"This paper proves a complete algebraic characterization of the observable second- and third-order moments of a linear non-Gaussian acyclic model (LiNGAM). For any directed acyclic graph $G$, a positive definite covariance matrix $S$ and a symmetric third-order moment tensor $T$ are realizable under $G$ if and only if, for every vertex $v$, a matrix built from the moments of $v$, its parents, and its non-descendants has rank exactly equal to the number of parents of $v$. The characterization reduces membership testing to explicit determinantal conditions, extends the previously known polytree case to arbitrary DAGs, and contains the Gaussian local Markov property as the special case where third moments vanish. This matters because it turns the causal question of whether data could have come from a given graph into a check of simple rank constraints, the same algebraic structure that already powers non-Gaussian causal discovery algorithms.","feed_headline":"Moment ranks reveal which causal DAG made your data","feed_subtitle":"Explicit rank conditions on second and third moments identify the full causal DAG, not just its equivalence class.","key_machinery":"The central object is the blocked moment matrix $M_v$, whose top rows are indexed by the parents of $v$, whose bottom row is indexed by $v$, and whose columns are indexed by non-descendants of $v$ for the covariance entries together with pairs $(\\text{non-descendant}, \\text{any vertex})$ for the flattened third-moment entries. The condition that $M_v$ has rank $|\\mathrm{pa}(v)|$ is equivalent to the bottom row lying in the span of the parent rows, and those span coefficients become the structural coefficients $\\lambda_{uv}$ of the LiNGAM. The same matrix simultaneously encodes the second-moment relations, namely the classical local Markov property recovered from the covariance block alone, and the third-moment relations that carry the additional directional information needed to identify the full DAG rather than its Markov equivalence class.","core_discovery":"Formally, Theorem 3.1 states that $(S,T)$ belongs to the second- and third-order moment model $M_{\\le 3}(G)$ if and only if each matrix $$M_v = \\begin{pmatrix} S_{\\mathrm{pa}(v),\\mathrm{nd}(v)} & T_{\\mathrm{pa}(v),\\mathrm{nd}(v)\\times V} \\\\ S_{v,\\mathrm{nd}(v)} & T_{v,\\mathrm{nd}(v)\\times V} \\end{pmatrix}$$ has rank $|\\mathrm{pa}(v)|$. The necessity direction comes from the structural equation $X_v = \\sum_{u\\in\\mathrm{pa}(v)} \\lambda_{uv} X_u + \\varepsilon_v$: because $\\varepsilon_v$ is independent of every non-descendant $X_w$, the vector $(-1,\\lambda_{\\mathrm{pa}(v),v})$ is a left null vector of $M_v$, and positive definiteness of $S$ prevents the rank from dropping below $|\\mathrm{pa}(v)|$. The sufficiency direction reads the regression coefficients $\\lambda_{uv}$ off the linear dependence of the bottom row of $M_v$ on the parent rows, and then shows that the same coefficients diagonalize both the transformed covariance matrix and the transformed third-order moment tensor, recovering the independent error terms. The model is therefore cut out inside the positive definite cone by the vanishing of all $(|\\mathrm{pa}(v)|+1)$-minors, giving a finite list of polynomial constraints that separate one DAG from another.","pith_inferences":["One natural next step the paper does not take is to turn the rank conditions into a model-selection score: among graphs that explain the covariance equally well, the graph whose $M_v$ ranks are closest to $|\\mathrm{pa}(v)|$ would be preferred, and finite-sample thresholds could be derived from the singular-value behavior the authors illustrate.","The authors' remark that higher-order cumulants should behave analogously suggests a concrete testable extension: replacing third moments by fourth cumulants should yield an analogous rank characterization for models with symmetric errors, where third moments vanish and the current theorem gives no grip.","The discussion of latent-variable LiNGAM hints that eliminating third-moment blocks from these rank conditions, rather than adding them, might produce new constraints for hidden-variable models, including some that cannot be expressed by covariance determinants alone.","The ROC-style threshold experiment in the paper suggests the rank tests have practical statistical behavior worth formalizing, but no asymptotic distribution of the condition numbers is proven here; that would be needed for principled threshold selection."],"forward_implications":["Membership of a moment pair $(S,T)$ in the LiNGAM model for any DAG is decided by finitely many rank conditions, so model selection between candidate DAGs can be framed as checking which graph's matrices $M_v$ drop to the required ranks.","The covariance-only part of the matrices recovers the Gaussian local Markov property, so the theorem is a genuine extension of classical covariance constraints rather than a separate theory.","The polytree characterization from earlier work follows as a special case, and the new proof shows the polytree $2\\times 2$ determinants are submatrices of the more general rank conditions.","Source and sink nodes can be detected by the same machinery: the rank-one matrices used in existing source-finding algorithms are a direct corollary, and a new sink-detection procedure based on the condition numbers of $M_v$ works in simulations.","Because the constraints are all minors of explicit matrices, they give concrete polynomial equations usable in goodness-of-fit tests for whether data follow any linear non-Gaussian acyclic model."],"supporting_citations":[{"why":"Gives the polytree characterization that Theorem 3.1 generalizes, and supplies the moment parametrization lemma used throughout the proof.","marker":"[ADG+23]"},{"why":"Provides the algebraic statistics background, the local Markov property rank constraints, and the LiNGAM parametrization that the new constraints extend.","marker":"[Sul18]"},{"why":"Develops the high-dimensional source-recovery algorithm whose 2-by-2 rank tests are recovered as a direct corollary of the present theorem.","marker":"[WD19]"},{"why":"Establishes trek separation for Gaussian graphical models, the classical rank-constraint theory that this paper extends to third moments.","marker":"[STD10]"},{"why":"Introduces the LiNGAM model and the identifiability of the full DAG from non-Gaussian data, which motivates the whole investigation.","marker":"[SHHK06]"},{"why":"Provides the multi-trek separation framework for linear structural equation models, the previous generalized separation tool that the present characterization completes.","marker":"[RS21]"},{"why":"Shows how algebraic constraints are used for goodness-of-fit tests for linear non-Gaussian models, an application the new rank conditions directly support.","marker":"[SD23]"}],"fun_headline_variants":["Rank conditions on moments pin down the causal graph exactly","Third moments break causal equivalence – full DAG recovery","Moment matrices reveal the exact causal DAG","Second and third moments uniquely identify the causal graph","Determinantal constraints: moment ranks select the true DAG"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The characterization stands or falls on the LiNGAM assumption that the noise variables are mutually independent with mean zero and finite third moments and that every common cause is included in the graph; if errors are dependent or a hidden confounder exists, the moment equations that force the matrices $M_v$ to drop rank no longer hold.","fun_headline_variants_meta":{"raw":{"variants":["Rank conditions on moments pin down the causal graph exactly","Third moments break causal equivalence – full DAG recovery","Moment matrices reveal the exact causal DAG","Second and third moments uniquely identify the causal graph","Determinantal constraints: moment ranks select the true DAG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000272,"raw_usage":{"total_tokens":1665,"prompt_tokens":1008,"completion_tokens":657,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":624,"completion_tokens_details":{"reasoning_tokens":581}},"tokens_in":624,"tokens_out":657,"duration_ms":6419,"temperature":1.0,"reasoning_tokens":581,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:49:23.349145+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The theorem would be refuted by a triple $(G,S,T)$ with $S$ positive definite and $T$ symmetric where every matrix $M_v$ has rank $|\\mathrm{pa}(v)|$ but $(S,T)$ is not in the image of the parametrization $S=(I-\\Lambda)^{-T}\\Omega^{(2)}(I-\\Lambda)^{-1}$, $T=\\Omega^{(3)}\\bullet (I-\\Lambda)^{-1}\\bullet (I-\\Lambda)^{-1}\\bullet (I-\\Lambda)^{-1}$ for any $\\Lambda\\in\\mathbb{R}^E$ and diagonal $\\Omega^{(2)},\\Omega^{(3)}$. For the three-vertex DAG $1\\to 2$, $1\\to 3$, $2\\to 3$, one can test this directly with a computer algebra system: the theorem predicts the model is cut out by the 2-minors of $M_2$ and 3-minors of $M_3$ saturated by the principal minors of $S$, so any point in that variety outside the parametrization image would disprove the sufficiency direction.","supporting_citations":[],"review_version":1}