{"id":"ea7333f8-b783-4022-9a2c-2fd2440e6241","arxiv_id":"2411.19556","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A differentiable VAE method with rank-Jacobian identifiability claims for nonlinear latent hierarchical models, undermined by a false core theorem and a circular proof.","lead":"This paper proposes a differentiable causal discovery algorithm for nonlinear latent hierarchical models, together with new identifiability theorems that aim to relax prior deterministic-function assumptions. The VAE-based method outperforms earlier latent-causal-discovery baselines on small synthetic graphs and on image transfer tasks, but the central identifiability theorem is flawed.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's formal rank–d-separation equality is false as stated: the proof derives rank = min(|Z|, |X|, |Y|), and Condition 1 permits minimal separators larger than the Jacobian's maximum rank.","rationale":"The reader's weakest_assumption correctly identifies the missing bound |Z| ≤ min(|X|, |Y|) as the load-bearing flaw: the proof only establishes a capped rank formula, not the stated equality. I would sharpen the failure slightly: the singleton-set example with a scalar Jacobian and a size-2 separator lies outside the theorem's 'r < |X|, |Y|' quantifier, so the formal Eq. (3) is the clearest place where the theorem is false. The same missing dimension bound is at issue, which is why this is partial rather than full disagreement. The concern is load-bearing because the identifiability theorem (Theorem 3) and Lemmas 1–3 all use the rank-separator link; if that link requires an unstated bound, the advertised identifiability result is unsupported as written. The paper does have genuine empirical content: the synthetic experiments show strong SHD/F1 performance on small trees and v-structures compared with the chosen baselines, and the image experiments demonstrate a scalable end-to-end pipeline. However, no code is provided, the synthetic graphs are small, and the image evaluations compare only against representation-learning baselines rather than other latent causal discovery methods, so the experiments cannot repair the theoretical gap. A correction of Theorem 1 to rank(J_f) = min(min_Z |Z|, |X|, |Y|) under a non-circular faithfulness argument might salvage the lemmas in the regimes where |Z| is genuinely small, but that correction is neither stated nor proved in the paper.","tokens_in":21224,"tokens_out":14351,"duration_ms":133705,"concrete_test":"Construct the four-latent counterexample described above with generic smooth structural equations (e.g., random-weight Tanh or LeakyReLU maps with independent noise) and estimate the 2×2 Jacobian J_f of f(x)=E[y|x] at several values of x. Verify by d-separation that the smallest latent set separating X={x1,x2} from Y={y1,y2} is {z1,z2,z3,z4} of size 4, while the estimated Jacobian has rank 2 (or at most 2). This directly disproves Eq. (3). In parallel, re-derive Proposition 1 and Theorem 1 without mutual citation; if the generic-rank claim cannot be established independently from Lemma 5, then Condition 2's faithfulness justification remains unproven.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 1 derives rank(J_f) ≤ min(|Z|, |X|, |Y|) and then, via Proposition 1, concludes rank(J_f) = min(|Z|, |X|, |Y|). The paper then reads this as rank(J_f) = |Z| in Eq. (3). That step is only licensed when |Z| ≤ min(|X|, |Y|). No such bound is assumed, and Condition 1 does not imply it: Condition 1 explicitly allows a measured variable with several latent parents (e.g., x12 in Figure 1), so the minimal d-separating set between two measured sets can exceed the maximum possible rank of the Jacobian. Concretely, take latent variables z1..z4 with measured children x1 (parents z1,z2), x2 (parents z3,z4), y1 (parents z1,z3), y2 (parents z2,z4), plus two extra pure children for each latent to satisfy Condition 1(i) and unit path lengths for Condition 1(ii). The smallest latent set d-separating X={x1,x2} from Y={y1,y2} has size 4, while J_f is 2×2 and hence has rank at most 2, so the formal equality rank(J_f) = min_Z |Z| fails. The restricted biconditional for r < |X|, |Y| may survive this particular example, but the theorem as stated and its proof are invalid; the later lemmas rely on Theorem 1, and Proposition 1's generic-rank argument is also circular because it invokes Theorem 1, while Theorem 1's proof invokes Proposition 1.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a theoretical framework for identifying nonlinear latent hierarchical causal models from observational data, claiming to relax the deterministic-function assumption of Kong et al. (2023). The central result, Theorem 1, states that the rank of the Jacobian of the conditional expectation E[y|x] equals the size of the smallest latent set that d-separates the measured sets X and Y. Building on this, the paper derives an identifiability theorem (Theorem 3) and presents a differentiable causal discovery algorithm based on a VAE with Gumbel-softmax masks and structural constraints. Experiments on synthetic graphs and image datasets are reported to support the method's accuracy and scalability.","tokens_in":21640,"tokens_out":11151,"duration_ms":87175,"significance":"If the main theorem were correct, the paper would be a substantial step forward: it would provide a continuous, scalable rank-based criterion for latent hierarchical structure without requiring linearity or deterministic invertibility, and the empirical work engages seriously with baselines and evaluation concerns. However, the central theorem is false as stated, and the proof chain contains a circular dependency between Proposition 1 and Theorem 1. Because the identifiability result and the algorithm's structural constraints both rest on Theorem 1, the paper's core theoretical contribution is not established. The empirical results, while suggestive, cannot compensate for the invalidity of the theory that motivates them.","major_comments":[{"comment":"The proof derives rank(J_f(x)) ≤ min(|Z|, |X|, |Y|) and then, after invoking Proposition 1, concludes rank(J_f(x)) = min(|Z|, |X|, |Y|) and subsequently equates this with |Z|. The final step is licensed only when |Z| ≤ min(|X|, |Y|). Condition 1 does not imply this bound: it explicitly allows a measured variable to have several latent parents (e.g., x12 in Figure 1), so the minimal d-separating set between two measured sets can be larger than either set. For example, take latent variables z1,...,z4 with measured children x1 (parents z1,z2), x2 (parents z3,z4), y1 (parents z1,z3), y2 (parents z2,z4), and add two pure children to each latent with unit path lengths to satisfy Condition 1. The smallest latent set d-separating X={x1,x2} from Y={y1,y2} has size 4, while J_f is a 2×2 matrix and hence has rank at most 2; for a generic faithful parameter choice its rank is 2, contradicting Eq. (3). This also breaks the claimed biconditional for r < |X|, |Y| when the minimal separator exceeds the Jacobian's maximal possible rank. Since Lemmas 1-3 and Theorem 3 all invoke Theorem 1, the identifiability theory is unsupported.","section":"§4, Theorem 1 and Appendix A.2"},{"comment":"The proof of Proposition 1 cites 'the proof of Theorem 1' for the decomposition J_h = J_f ∘ J_g, while the proof of Theorem 1 invokes Proposition 1 to conclude that the Jacobian attains its maximal rank almost everywhere. This is a direct circular dependency between the two results. In addition, Proposition 1's application of Lemma 5 requires |X| ≥ |Z| and |Y| ≥ |Z|; no such dimension condition is stated in Condition 1 or Condition 3, and the example in the previous comment shows that it can fail. The generic-rank claim is the bridge from the inequality rank(J_f) ≤ min(|Z|, |X|, |Y|) to the equality rank(J_f) = min(|Z|, |X|, |Y|), so this gap is load-bearing.","section":"Appendix A.1 and Appendix A.2"},{"comment":"The theorem's quantifier restricts attention to r < |X|, |Y|, but the formal display, Eq. (3) (and Eq. (11) in the proof), asserts an unqualified equality rank(J_f) = min_Z |Z|. The proof treats the unqualified equality as its target and then derives it by the invalid step identified above. As written, the formal equality is false for models where the minimal separator is larger than both |X| and |Y|. The restricted biconditional, even if it could be repaired under additional assumptions, is not sufficient for the subsequent lemmas and Theorem 3, which apply the equality to arbitrary measured sets without verifying the dimension bound.","section":"§4, Theorem 1 statement and Eq. (3)"},{"comment":"Theorem 3 assumes an oracle r(S,T) that outputs the minimum number of latent variables d-separating two measured sets, and the only route proposed in the paper for obtaining this oracle is the Jacobian rank criterion of Theorem 1. Since Theorem 1 is invalid, Theorem 3 does not establish identifiability from the observed distribution. Furthermore, the proof of Lemma 2 in Appendix A.5 asserts without derivation that any set d-separating S∪{c} from T must contain P and that r(S,T) = |P|; this presupposes the identification problem that the lemma is meant to solve. Thus the paper's central identifiability claim is not supported.","section":"§4, Theorem 3 and Appendix A.7"}],"minor_comments":[{"comment":"In condition (1) of Lemma 2, the right-hand side reads '= P' where P is a set; it should presumably be '= |P|'.","section":"Appendix A.5, Lemma 2"},{"comment":"The constraint expression in Eq. (8) is malformed: it should read ∥Mi,:∥1 (∥Mi,: ⊙ ∏_{j≠i}(1−Mj,:)∥1 − 2) ≥ 0 for each i.","section":"Section 5.2, Eq. (8)"},{"comment":"The figure caption lists 'z8' twice in the sequence of latent variables; the second occurrence is likely a typo and should be a different latent variable.","section":"Figure 1 caption"},{"comment":"There are typos, including 'signficant' and 'devation', and the explanation of the SHD-over-permutations evaluation is given twice and could be condensed.","section":"Appendix B.1"}],"recommendation":"reject","confidential_remarks":"The core theoretical claim is false, and the flaw is not merely presentational: the mismatch between the Jacobian rank and the size of the minimal d-separating set is intrinsic to models with non-pure children having multiple parents, which the paper explicitly aims to allow. A repair would require either a different observable criterion or a substantial restriction of the model class, changing the paper's scope. I therefore recommend rejection rather than major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the central theoretical claim is not correct, and the proof as written is circular. Theorem 1 asserts that rank(J_f) equals the size of the smallest latent d-separating set, but the proof only establishes rank = min(|Z|, |X|, |Y|). That gap is not cosmetic: under the paper's own Condition 1, a measured variable can have two latent parents, so the minimal separator between two measured sets can be strictly larger than the Jacobian's maximum rank. The stress-test note's four-latent example is a concrete counterexample to the theorem as stated. So the identifiability theory, and everything that depends on it (Lemmas 1–3, Theorem 3), does not go through.\n\nCredit where due: the paper does something genuinely new on the algorithmic side. Combining a VAE with Gumbel-softmax relaxation to learn a latent hierarchical graph in one pass, rather than training O(l n^2) generative models as in Kong et al., is a reasonable idea, and the synthetic experiments show it can recover small ground-truth graphs more accurately than the baselines tested. The image experiments are exploratory but not misleading: the authors clearly label them as qualitative, and the CMNIST/CelebA transfer results are suggestive. The paper is also honest in the appendix that the learned MNIST graph does not exactly satisfy Condition 1.\n\nThe soft spots beyond the false theorem: the proof of Proposition 1 invokes Theorem 1, and Theorem 1's proof invokes Proposition 1; that is a circular dependency, and the dimension condition in Lemma 5 (m,n ≥ p) is exactly the condition that the counterexample violates. Lemma 1's proof also uses the same faulty rank-separation equivalence, so the pure-child identification criterion is unsupported. The paper gives no code, and the synthetic evaluation uses only four small graphs with three runs each, so the empirical claim is thin. The image experiments do not compare against other latent causal discovery methods (e.g., Kong et al.), only against representation learning baselines.\n\nWho this is for: someone interested in differentiable structure learning for latent hierarchical models might find the algorithm worth trying, but the paper's theoretical contribution cannot be cited as a proof. I would not publish as is. If the authors can fix the theorem by adding the condition |Z| ≤ min(|X|,|Y|) (which may be impossible under Condition 1) or by redefining what is identifiable, the algorithm could be a useful addition. As it stands, the load-bearing result is false.\n\nRecommendation: I'd send it to peer review rather than desk-reject, because the algorithmic idea is novel and the error is instructive; a good referee can point to the specific gap and give the authors a chance to either repair the theory or honestly restrict the claims. But my own verdict would be reject unless the theory is fixed.","headline":"The paper's central theorem is false as stated, and the proof is circular; the algorithmic contribution is real but the theory cannot be used.","tokens_in":22129,"tokens_out":4060,"would_cite":false,"duration_ms":32990,"reading_group":"no","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 nonlinear latent hierarchical causal models are identifiable from observational data without assuming latent variables or noise are deterministic functions of measurements, and obtains the graph via a Jacobian rank…","keywords":["latent hierarchical causal models","causal discovery","identifiability","Jacobian rank","nonlinear SEM","variational autoencoder","differentiable optimization","d-separation"],"falsifier":"Simulate a nonlinear hierarchical model satisfying the paper's Condition 1 with a measured variable that has two latent parents (as in the paper's Figure 1), and compute the rank of the Jacobian of E[Y|X] for X equal to that single variable and Y any other measured set. The minimal d-separating set has size 2, while the scalar-input Jacobian has rank at most 1; if the paper's lemmas rely on these two numbers being equal, they are falsified on this graph.","tokens_in":21022,"feed_emoji":"🧩","tokens_out":16309,"duration_ms":124500,"temperature":0.7,"pith_summary":"This paper tries to show that the causal structure of nonlinear hierarchical models with unobserved variables can be recovered from purely observational data, and that the earlier deterministic-function assumption (latents and noise are invertible functions of the measurements) can be dropped. The load-bearing criterion is a new rank test: the rank of the Jacobian of the conditional expectation E[y|x] equals the number of latent variables in the smallest set that d-separates x from y. From this test the paper proves the whole graph is identifiable up to relabeling of the latents, and it builds a differentiable algorithm—a VAE with Gumbel-softmax adjacency masks and a two-pure-children constraint—that learns the graph in one continuous optimization run. The method is demonstrated on synthetic graphs and on image data, where the learned hierarchy is interpretable and transfers across distribution shifts.","feed_headline":"Derivative rank recovers latent causal structure","feed_subtitle":"A Jacobian rank test plus a trainable VAE recovers nonlinear hierarchical causal structure without deterministic-function assumptions.","key_machinery":"The load-bearing object is the Jacobian rank equality rank(∂E[y|x]/∂x) = size of the smallest d-separating latent set (Theorem 1). It converts a graphical notion, d-separation, into a continuous, data-observable quantity, which is what allows discrete structure search to be replaced by differentiable optimization. Supporting machinery: Theorem 2 lets measured pure descendants stand in for latent sets; Lemmas 1–3 turn rank equalities into rules for identifying pure-child groups, the parent set of non-pure children, and parentless measured variables; and the algorithm uses a block upper-triangular Gumbel-softmax adjacency matrix inside a VAE whose decoder respects the SEM, with a constraint that each latent row has at least two pure children.","core_discovery":"The central claim is Theorem 1: under a generalized faithfulness condition and a differentiability condition, for any two measured sets X and Y with r < |X|, |Y|, the rank of J_f = ∂E[y|x]/∂x equals the size of the smallest set of latent variables that d-separates X from Y. This generalizes the rank-deficiency constraints used in linear latent-variable models to the nonlinear case and removes the deterministic-function requirement of Kong et al. (2023). The paper further shows that pure measured descendants can stand in for latent sets (Theorem 2), which allows a recursive procedure: identify the lowest latent layer from the measured variables using three rank-based lemmas, then lift to the next layer using the identified pure children as surrogates. The result (Theorem 3) is that the binary adjacency matrix M of the hierarchical graph is identifiable up to permutation of the latent variables.","pith_inferences":["The Jacobian-rank criterion could likely be repurposed as a nonparametric test for latent confounding between arbitrary variable sets, whenever each side has enough measured variables to carry the dimension.","The two-pure-children condition is probably not the exact identifiability boundary; nonlinear mechanisms with sparse footprints may still permit identification under additional smoothness or noise assumptions.","The need to fix the number of latent layers in advance suggests a practical extension: run the differentiable search with varying layer counts and select by held-out likelihood, effectively performing structure selection over depth.","Because the rank test is based on conditional expectations, finite-sample estimation error in the regression function will bias rank estimates downward; a bias-correction or a hypothesis-testing version of the rank equality would be a natural next step."],"forward_implications":["Nonlinear latent hierarchical causal models become identifiable from observational data without requiring latent variables or noise to be deterministic functions of the measurements.","The Jacobian rank of E[y|x] provides a general, distribution-level test for how many hidden variables separate any two measured sets, extending rank constraints beyond linear models.","The differentiable VAE-based algorithm learns the graph in one continuous optimization run, avoiding the iterative training of many generative models and scaling to hundreds of latent variables.","On image data, the recovered hierarchy is interpretable and its representations transfer across distribution shifts (CMNIST, CelebA), supporting the use of causal latent structure for domain adaptation."],"supporting_citations":[{"why":"Prior identifiability result for nonlinear latent hierarchical models under the deterministic-function assumption (z,ε)=f(x); the paper's Theorem 3 removes that assumption.","marker":"Kong et al. (2023)"},{"why":"Linear latent hierarchical identifiability via rank constraints on covariance; the paper's Jacobian rank test is the nonlinear generalization of this idea.","marker":"Huang et al. (2022)"},{"why":"Provides the d-separation and faithfulness framework used to relate conditional independence to graph separation.","marker":"Spirtes et al. (2001)"},{"why":"Introduces the pure-children requirement for linear latent variable models, which Condition 1 adapts to the nonlinear hierarchical setting.","marker":"Silva et al. (2006)"},{"why":"NOTEARS reformulates discrete graph search as continuous optimization with algebraic acyclicity constraints, the starting point for the paper's differentiable discovery approach.","marker":"Zheng et al. (2018)"}],"fun_headline_variants":["Jacobian rank exposes latent causal hierarchy","Differentiable search for latent causal models","Latent hierarchy from Jacobian rank test","Causal discovery without deterministic latents","Scalable latent causal structure learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes the smallest set of hidden variables that separates two measured groups is never larger than the smaller of the two groups, yet the paper's own conditions allow a single measured variable to have two latent parents, where that assumption fails for singleton groups.","fun_headline_variants_meta":{"raw":{"variants":["Jacobian rank exposes latent causal hierarchy","Differentiable search for latent causal models","Latent hierarchy from Jacobian rank test","Causal discovery without deterministic latents","Scalable latent causal structure learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000264,"raw_usage":{"total_tokens":1566,"prompt_tokens":872,"completion_tokens":694,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":488,"completion_tokens_details":{"reasoning_tokens":633}},"tokens_in":488,"tokens_out":694,"duration_ms":6762,"temperature":1.0,"reasoning_tokens":633,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:08:24.585024+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a nonlinear hierarchical model satisfying the paper's Condition 1 with a measured variable that has two latent parents (as in the paper's Figure 1), and compute the rank of the Jacobian of E[Y|X] for X equal to that single variable and Y any other measured set. The minimal d-separating set has size 2, while the scalar-input Jacobian has rank at most 1; if the paper's lemmas rely on these two numbers being equal, they are falsified on this graph.","supporting_citations":[{"cited_title":"Learning the structure of linear latent variable models","cited_arxiv_id":null,"evidence_quote":"Introduces the pure-children requirement for linear latent variable models, which Condition 1 adapts to the nonlinear hierarchical setting."}],"review_version":1}