{"id":"3db41af7-9e7f-4868-9a46-71a65ffb0066","arxiv_id":"2501.10897","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Each latent variable with at least two pure observed children makes the bipartite graph identifiable from the population tensor via rank tests on unfolded matrices.","lead":"This paper proves that the connection graph between hidden and observed variables in discrete latent bipartite models, such as Noisy-Or networks and restricted Boltzmann machines, can be recovered from the exact joint distribution when each hidden variable has at least two pure observed variables attached only to it. The proof is constructive and uses tensor unfolding, and it also yields the number of hidden variables.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1's rank certificate fails for observed variables with no latent neighbours: an isolated Yj paired with a pure variable gives rank ≤ H, violating the 'only if' and the constructive recovery in Theorem 1.","rationale":"The reader's weakest_assumption mentions the omitted co(j)=∅ case as secondary to the generic-rank gap. The stress test shows the empty-co(j) case is actually the most load-bearing flaw: it is a structural configuration explicitly allowed by G* arbitrary, and it makes the rank certificate fail for every parameter value, including generic ones. Thus adding a generic qualifier to Proposition 1 cannot repair the theorem. The concrete 3-variable example settles the matter by exhibiting a valid instance where rank([T]_{1,3},:) = H even though co(1)={1} and co(3)=∅. The constructive argument appears salvageable if the authors add the mild assumption that no observed variable is isolated from the latent layer, or if they first screen out variables independent of all others before running the rank tests. A generic-parameters qualifier is also needed for the separate rank-certificate gap in Lemma 3 and §A.1. With these amendments, the core tensor-unfolding idea remains plausible and novel; without them, Theorem 1 as stated is false. The verdict therefore remains CONDITIONAL rather than UNCHANGED, because the theorem needs explicit correction.","tokens_in":29758,"tokens_out":18446,"duration_ms":202042,"concrete_test":"Set V=H=2, K=1, J=3. Let A1~Bernoulli(1/2); Y1,Y2 be pure children of A1 with full-rank CPTs, e.g. P(Y=0|A1=0)=0.9, P(Y=0|A1=1)=0.1; let Y3 have co(3)=∅, e.g. Y3~Bernoulli(1/2) independent of everything. This satisfies Assumptions 1–2 and the row-permuted G=(1;1;0), so Theorem 1 applies. Compute the 4×2 unfolding [T]_{1,3},:; its entries are P(Y3=y3)P(Y1=y1,Y2=y2), so the matrix has two identical row blocks and rank = rank(P(Y1,Y2)) = 2 = H. Thus rank([T]_{1,3},:)≤H while co(1)={1} and co(3)=∅, directly contradicting Proposition 1's 'only if' and disproving the constructive identifiability claim.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Proposition 1 asserts, for every j1≠j2, rank([T]_{j1,j2},:)≤H iff co(j1)=co(j2)={k}. The 'only if' is false when one of the two observed variables has no latent connections, a case allowed by Theorem 1 since G=(I_K;I_K;G*) has G* arbitrary. Take co(j1)=∅ and co(j2)={k}. Local independence makes Y_{j1} independent of all other variables, so [T]_{j1,j2},: = P(Y_{j1}) ⊗ P(Y_{j2},Y_rest) up to row/column grouping, and its rank equals rank(P(Y_{j2},Y_rest)), which is at most H because Y_{j2} depends on the rest only through the H-state latent A_k. Hence rank≤H although co(j1) and co(j2) are not the same singleton. The algorithm would classify the isolated variable as a pure child of A_k and can output a wrong graph. This is a structural gap, not a generic-parameter exceptional set; it persists for all parameters, including generic ones. A separate generic-only issue appears in the proof of the 'only if' (Lemma 3 and 'for generic parameters' in §A.1), but the isolated-variable counterexample already falsifies the stated theorem. The abstract identifiability of G may survive by other means, but the claimed constructive recovery via these rank certificates does not.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a constructive identifiability result for the bipartite loading graph in discrete latent bipartite graphical models, covering directed models such as Noisy-Or Bayesian networks and undirected models such as restricted Boltzmann machines. The central claim is that, under a sparsity condition in which every latent variable has at least two pure observed children/neighbors, the population tensor can be unfolded into matrices whose ranks certify which observed variables are connected to which latent variables. The proof gives a population-level algorithm recovering both the number of latent variables and the adjacency matrix up to column permutation. The main theorem is supported by Proposition 1, a rank test for pure (single-neighbor) structure, and Proposition 2, a rank test for multi-neighbor structure; a side proposition addresses identifiability of the remaining continuous parameters.","tokens_in":30105,"tokens_out":7244,"duration_ms":79439,"significance":"If the main theorem were correct as stated, it would be a useful contribution: it extends graph identifiability beyond linear Gaussian factor models, gives a constructive population-level algorithm rather than a nonconstructive tensor-decomposition argument, and covers flexible nonlinear discrete models with cycles. The tensor-unfolding perspective and the use of pure observed variables are clear and well motivated. The paper also correctly identifies the limitation of Kruskal-based proofs, which usually require knowing the number of latent variables in advance. However, the main result is currently overstated: Proposition 1 is false as stated for observed variables with no latent connections, and the strict rank inequalities in the proofs are established only generically while Theorem 1 is stated without a generic qualifier. These are load-bearing issues, but they are local in the sense that adding a no-isolated-variable condition and a generic-parameter qualifier may repair the main claim.","major_comments":[{"comment":"Proposition 1's 'only if' direction is false for observed variables with no latent connections, a case allowed by Theorem 1 because G* is an arbitrary binary matrix and may contain all-zero rows. Take co(j1)=∅ and co(j2)={k}. Local independence gives [T]_{j1,j2},: = P(Y_{j1}) ⊗ P(Y_{j2},Y_rest) up to row/column permutation, so its rank equals rank(P(Y_{j2},Y_rest)). Since Y_{j2} depends on the rest only through the H-state latent A_k, rank(P(Y_{j2},Y_rest)) ≤ H. Thus rank([T]_{j1,j2},:) ≤ H although co(j1) and co(j2) are not the same singleton. The constructive algorithm in Theorem 1 can therefore group an isolated observed variable together with the pure children of any latent variable and output a wrong graph. The proof in §A.1 skips exactly this case when it asserts that failure of the singleton condition implies |co(j1)∪co(j2)| ≥ 2. The theorem needs an explicit assumption excluding empty co(j) (for example, every row of G has at least one 1), and the empty case needs to be analyzed rather than implicitly excluded.","section":"§4.1, Proposition 1 and Theorem 1"},{"comment":"Propositions 1 and 2 are stated as deterministic equivalences under Assumptions 1 and 2, but their proofs only establish the strict rank inequalities for generic parameters. The 'only if' direction of Proposition 1 invokes Lemma 3, which is itself stated 'for generic parameters', and additionally uses generic full rank of P2 and P3 in equation (18). Proposition 2 similarly uses generic full rank of P2 and P3 in equation (19). Consequently, for parameter values lying on the excluded algebraic variety, the strict inequalities rank > H and rank > H^{K-1} can fail, and the rank certificates no longer separate the graph structures. Theorem 1 therefore needs either a generic-parameter qualifier in its statement or additional conditions that rule out the exceptional variety. This is a separate issue from the empty-neighborhood counterexample, and both must be fixed before the main claim is true.","section":"§A.1 and §A.2, generic qualifier"},{"comment":"The proof of the 'if' direction of Proposition 2 invokes Lemma 3 to lower-bound rank(P1,2), but P1,2 is a conditional probability table for |co(j)| observed variables, not a two-variable unfolding. The lower bound rank(P1,2) > H^{|co(S1)|} appears to follow from Lemma 2 with S1=co(j)\\setminus{k} and S2={j}, not from Lemma 3. This is likely a citation error, but it should be corrected because the current reference makes the proof difficult to verify.","section":"§A.2, proof of Proposition 2"}],"minor_comments":[{"comment":"The word 'bipartige' is a typo for 'bipartite' in the discussion of cognitive diagnostic models.","section":"§2.3"},{"comment":"The sentence 'For a set S, denotes its cardinality by |S|' is missing a subject; it should read 'For a set S, we denote its cardinality by |S|'.","section":"§3"},{"comment":"The statement relies on Theorem 3.3 of Gu (2024), cited as 'Bernoulli, page to appear'. If that reference is still in press, the paper should make the dependence explicit in the main text so readers can assess the side result.","section":"§4.3, Proposition 3(a)"},{"comment":"In the proof, the sentence '1^T E_{β_{h1},r} = 1^T E_{β_{h1},r} = 0' should more precisely say that both E_{β_{h1},r} and E_{β_{h2},r} are conditional probability vectors, so each sums to 1 and their difference sums to 0; the displayed equality as written appears to contain a typo.","section":"§B.2, Lemma 2 proof"}],"recommendation":"major_revision","confidential_remarks":"The paper has a nice tensor-unfolding idea and the conditional contribution is potentially publishable, but the main theorem is false as stated because of the empty-neighborhood case and the missing generic qualifier. I would encourage the editor to ask for a careful revision that states the theorem with the necessary qualifications and verifies whether the constructive algorithm survives under those qualifications. The reliance on Gu (2024) for Proposition 3(a) should also be made explicit and checked before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the supplementary proof of Proposition 1 before trusting Theorem 1. The constructive rank certificate is genuinely nice, but the theorem is over-stated: for observed variables with no latent neighbours, the 'only if' in Proposition 1 fails, so the claimed recovery algorithm can pair an isolated variable with a pure child of some latent and output a graph with spurious edges.\n\nWhat is new and good: the idea of using pairwise tensor unfoldings to certify pure children/neighbours, and then using rank inequalities for multi-parent structures, is a real extension of the quartet tests in latent tree models and of Kruskal-based parameter identifiability. The factorization in Section 4.2 is illuminating, and the result gracefully covers directed and undirected models under local independence. The paper is well written, and the proofs are mostly standard.\n\nThe main problem is Proposition 1. The proof of the 'only if' direction only handles |co(j1) ∪ co(j2)| ≥ 2. But G* in Theorem 1 is an arbitrary binary matrix, so zero rows are allowed; co(j) = ∅ is in scope. For such a j, Y_j is independent of everything else. The unfolding [T]_{j1,j2},: with any other observed variable j2 has rank at most H (indeed rank 1 times the rank of a matrix involving a pure child of a single latent), while co(j1) and co(j2) do not share a single latent. So rank ≤ H even though the two variables are not both pure children of the same latent. This is a structural counterexample; it persists for all parameter values, not just a measure-zero set. It breaks the constructive part of Theorem 1 as written.\n\nSeparately, the proofs of Propositions 1 and 2 conclude strict rank inequalities only 'for generic parameters', but Theorem 1 is stated without that qualifier. The generic caveat may be acceptable for identifiability of the graph, but for a constructive algorithm that takes the population tensor as input, the claim that the graph can 'uniquely be recovered' needs to say what happens on the exceptional set. This is softer than the isolated-variable counterexample, but both need addressing.\n\nThe core factorization idea is sound and worth taking seriously. With an added non-isolation assumption (or a pre-processing step that removes variables independent of all others) and a generic-parameter qualifier, the main result can likely be repaired. In the current form, Theorem 1 is false.\n\nThis paper is for algebraic statisticians working on discrete latent variable models and for people doing population-level structure learning for RBMs/Noisy-OR networks. It deserves a serious referee because the method is novel and the proof strategy is reusable, but the revision should be major. I would send it to review and make the isolated-variable issue the first thing the referees check.","headline":"The tensor-unfolding certificate is a real idea, but Theorem 1 is false as stated: isolated observed variables break Proposition 1's rank test.","tokens_in":30567,"tokens_out":3604,"would_cite":false,"duration_ms":37141,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that in discrete latent bipartite graphical models, the bipartite graph between observed and latent variables is identifiable—up to permutation of the latent variables—from the joint distribution of the observed…","keywords":["Algebraic Statistics","Discrete Graphical Model","Identifiability","Latent Variable Model","Noisy-Or Bayesian Network","Restricted Boltzmann Machine","Structure Learning","Tensor Unfolding"],"falsifier":"A concrete falsifier of the unconditional statement: choose a model satisfying Assumptions 1 and 2 with each latent having exactly two pure observed variables, but pick the conditional probability tables so that $P(Y_{\\{j_1,j_2\\}}\\mid A_{co(j_1)\\cup co(j_2)})$ has rank no greater than $H$ for a pair $j_1,j_2$ that do not share a single pure latent. Then the unfolding $[T]_{\\{j_1,j_2\\},:}$ would have rank at most $H$, violating the 'only if' direction of Proposition 1, so two distinct graphs would produce the same population tensor and Theorem 1's conclusion would fail.","tokens_in":29566,"feed_emoji":"🧩","tokens_out":9519,"duration_ms":84079,"temperature":0.7,"pith_summary":"This paper proves a new identifiability result for discrete latent bipartite graphical models, a family that includes Noisy-Or Bayesian networks for medical diagnosis and restricted Boltzmann machines. Its central claim is that the bipartite graph G, which records which observed variables load on which latent variables, is uniquely determined by the joint distribution of the observed variables, up to relabeling the latent variables, under a simple sparsity condition: each latent variable must have at least two 'pure' observed variables connected only to it. The proof is constructive: by reshaping the J-way probability tensor into matrices in specific ways and reading off matrix ranks, one recovers both the number of latent variables K and the full graph. Because the condition does not require linearity or a specific directed or undirected structure, the result covers a wide range of nonlinear and interaction-effect models. A consequence is that the substantively interpretable part of the model—the graph—can be identified even when the full set of continuous parameters may not be.","feed_headline":"Two pure measurements suffice to identify latent graphs","feed_subtitle":"A constructive proof shows the observable distribution fixes both the number of latents and their connections, with no linearity assumed.","key_machinery":"The central object is the $J$-way population probability tensor $T$, together with its unfoldings into matrices by grouping modes. The carrying mechanism is the factorization of each unfolded matrix through the latent variables: when $co(j_1)=co(j_2)=\\{k\\}$, the matrix $[T]_{\\{j_1,j_2\\},:}$ factors as $P(Y_{\\{j_1,j_2\\}}\\mid A_k)\\,P(A_k,A)\\,P(Y_{\\setminus\\{j_1,j_2\\}}\\mid A)^\\top$, forcing its rank at most $H$; the converse direction uses a rank lower bound on the conditional table $P(Y_{\\{j_1,j_2\\}}\\mid A_{co(j_1)\\cup co(j_2)})$ to show the rank exceeds $H$ generically. Proposition 2 uses the analogous factorization with the $2K$ already-identified pure variables as the column group, giving the threshold $H^{K-1}$. These rank certificates are what make the proof constructive.","core_discovery":"Theorem 1 states that under Assumptions 1 and 2, if the $J\\times K$ graph matrix $G$ contains two disjoint copies of the identity matrix $I_K$ after a row permutation—equivalently, $G=(I_K; I_K; G^\\star)^\\top$ for an arbitrary binary matrix $G^\\star$—then the bipartite graph is identifiable from the $J$-way population tensor $T$ with entries $T_{i_1,\\dots,i_J}=P(Y_1=i_1,\\dots,Y_J=i_J)$, up to a permutation of the $K$ latent columns. Identification is constructive: unfolding $T$ into matrices and testing their ranks yields both $K$ and the full incidence matrix. Specifically, for any pair of observed variables $j_1\\neq j_2$, the rank of the unfolding $[T]_{\\{j_1,j_2\\},:}$ is at most $H$ (the number of latent categories) exactly when $j_1$ and $j_2$ are both pure children of the same latent variable (Proposition 1); for a multi-parent observed variable $j$, the rank of the unfolding $[T]_{([K]\\setminus\\{k\\})\\cup\\{j\\},\\{K+1,\\dots,2K\\}}$ exceeds $H^{K-1}$ exactly when $k\\in co(j)$ (Proposition 2). These two rank certificates give a population-level structure-learning algorithm that the paper presents as the proof of Theorem 1.","pith_inferences":["The rank-threshold structure hints at a practical finite-sample estimator: replace the population tensor by the empirical contingency table and threshold the estimated singular values of the unfolded matrices; the paper does not analyze the sample regime, so finite-sample guarantees remain open.","The two-pure-measurement condition is analogous to anchor-item assumptions in psychometric measurement models, so the result may transfer to cognitive diagnosis and educational testing practice where such pure questions are already required by design.","Because the proof only uses rank certificates from unfoldings, the approach might extend to multilayer latent models or to graphs with more than two observed layers, though the paper does not pursue this.","For restricted Boltzmann machines, the result suggests that sparse loading structures are identifiable in a way that dense ones are not; this could motivate architectural constraints for interpretable deep generative models."],"forward_implications":["If Theorem 1 is correct, the exact population distribution determines both the number of latent variables and the bipartite graph up to latent label permutation, with no prior knowledge of $K$ required.","The identification holds across directed models like Noisy-Or networks and undirected models like restricted Boltzmann machines, with independent or dependent latents, and with nonlinear link functions and interaction effects.","The condition is interpretable and actionable: each latent trait needs at least two pure observed indicators, while all other observed variables may have arbitrarily dense connections.","Corollary 1 shows that only marginal tensors of order $2K$ are needed to identify the graph, so the full $J$-way contingency table is not required.","Proposition 3 adds that two pure children per latent identify the graph but not necessarily all continuous parameters, whereas three pure children per latent identify every parameter."],"supporting_citations":[{"why":"Supplies the generic-parameter framework and the Kronecker-product rank lemma used inside Lemma 3, and its Kruskal-based identifiability framework is the main comparison point for the paper's constructive approach.","marker":"Allman et al. (2009)"},{"why":"Provides Lemma 4, the Khatri-Rao product rank inequality used to establish the rank certificates in Propositions 1 and 2.","marker":"Stegeman and Sidiropoulos (2007)"},{"why":"The classical three-way tensor decomposition uniqueness theorem the paper contrasts with its constructive unfolding method and invokes in Proposition 3(b) for full parameter identifiability.","marker":"Kruskal (1977)"}],"fun_headline_variants":["Rank tests decode bipartite graph from observable tensor","Pure parents: two observed nodes identify each latent","Constructive proof identifies latent bipartite graph from tensor","Tensor unfolding shows two pure measurements fix latent graph","No linearity required: graph identifiable from pure pairs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The rank certificates that drive the construction are proved only for generic parameter values, while Theorem 1 is stated without the generic qualifier; if the true conditional probability tables lie on the excluded algebraic variety, the strict rank inequalities can fail and the graph need not be recoverable by the stated tests.","fun_headline_variants_meta":{"raw":{"variants":["Rank tests decode bipartite graph from observable tensor","Pure parents: two observed nodes identify each latent","Constructive proof identifies latent bipartite graph from tensor","Tensor unfolding shows two pure measurements fix latent graph","No linearity required: graph identifiable from pure pairs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000652,"raw_usage":{"total_tokens":3055,"prompt_tokens":1077,"completion_tokens":1978,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":693,"completion_tokens_details":{"reasoning_tokens":1915}},"tokens_in":693,"tokens_out":1978,"duration_ms":14655,"temperature":1.0,"reasoning_tokens":1915,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:52:12.553588+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete falsifier of the unconditional statement: choose a model satisfying Assumptions 1 and 2 with each latent having exactly two pure observed variables, but pick the conditional probability tables so that $P(Y_{\\{j_1,j_2\\}}\\mid A_{co(j_1)\\cup co(j_2)})$ has rank no greater than $H$ for a pair $j_1,j_2$ that do not share a single pure latent. Then the unfolding $[T]_{\\{j_1,j_2\\},:}$ would have rank at most $H$, violating the 'only if' direction of Proposition 1, so two distinct graphs would produce the same population tensor and Theorem 1's conclusion would fail.","supporting_citations":[{"cited_title":"S., Matias, C., and Rhodes, J","cited_arxiv_id":null,"evidence_quote":"Supplies the generic-parameter framework and the Kronecker-product rank lemma used inside Lemma 3, and its Kruskal-based identifiability framework is the main comparison point for the paper's constructive approach."},{"cited_title":"and Sidiropoulos, N","cited_arxiv_id":null,"evidence_quote":"Provides Lemma 4, the Khatri-Rao product rank inequality used to establish the rank certificates in Propositions 1 and 2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The classical three-way tensor decomposition uniqueness theorem the paper contrasts with its constructive unfolding method and invokes in Proposition 3(b) for full parameter identifiability."}],"review_version":1}