{"id":"731360ed-f897-4394-88a5-245637eabe80","arxiv_id":"2501.11622","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A kernel clustering method with a u-centered sample mapping is proposed to discover heterogeneous subgroups with different causal structures.","lead":"This paper introduces a clustering method that groups data samples by their underlying cause-and-effect relationships instead of by their values. It could help doctors, economists, and climate scientists find subgroups that react differently to the same cause.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5.1's proof reverses the sign rule of Theorem 4.1 (positive Φ means dependence, not independence) and conflates m-connectivity with pairwise dependence; the claimed isomorphism, and with it the guarantee that clusters are causally meaningful, is unsupported.","rationale":"The reader's weakest assumption is the sign rule in Theorem 4.1, and I agree that it is unjustified. However, the more load-bearing problem is that Theorem 5.1—the paper's strongest claim—is internally inconsistent with Theorem 4.1. Even if one granted the sign rule, the proof of Theorem 5.1 reverses it and identifies pairwise dependence with graphical m-connectivity, which are different concepts (faithfulness notwithstanding). The claimed isomorphism is therefore not merely missing a distributional justification; it is logically unsound. This directly undermines the central claim that clusters formed in the kernel space correspond to causal structure, and it cannot be repaired by a small technical fix. The experimental results do not compensate because the IOD prediction task uses the sign of a sum of kernel values, not the claimed isomorphism, and the Boston experiment relies on regression coefficients within arbitrary clusters. For these reasons, the reject verdict is appropriate and no change to the reader's verdict is needed.","tokens_in":17086,"tokens_out":6880,"duration_ms":70692,"concrete_test":"Generate n=200 samples from a collider DAG X1→X3←X2 with nonlinear additive noise, e.g., X1,X2~U(-1,1), X3 = X1^2 + X2^2 + N(0,0.1). Compute the sample mapping function Φ(S_i) from Definition 4.2 and the sign matrix sign(Φ). Compute the m-connectivity matrix I for m=2 based on Definition 5.1. If Theorem 4.1 holds, Φ_{1,2}<0 for most samples, giving sign(Φ)_{1,2}=-1, while I_{1,2}=+1 because X1 and X2 are connected via X3. A systematic mismatch between sign(Φ) and I refutes the claimed isomorphism. Alternatively, inspect the proof text itself: under Theorem 4.1, the sentence 'Φ>0 indicates X_j⊥⊥X_j′' in Theorem 5.1's proof is false, so no valid derivation of the isomorphism exists.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central guarantee is Theorem 5.1, which claims an isomorphism between the causal graph space and the causal matrix space. The proof is internally inconsistent. It states: 'If the corresponding matrix element is positive: Φ(S_i)_{j,j′} > 0, it indicates a causal relationship between the features X_j and X_j′: X_j ⊥⊥ X_j′.' But Theorem 4.1 proved the opposite: a positive aggregate Φ entry implies X_j and X_j′ are not independent. Thus the sign convention is reversed, so the proof cannot establish a coherent mapping. Independently, the proof conflates pairwise dependence with m-connectivity: N_G records the presence of a path of length at most m, whereas Φ encodes marginal dependence. In a collider X1→X3←X2, X1 and X2 are marginally independent (so Φ_{1,2} < 0, if Theorem 4.1 is accepted), yet they are connected by the path X1–X3–X2 and would be included in N_G for m=2. Hence the indicator matrices I and sign(Φ) cannot agree in general. Finally, the theorem is ill-typed: Φ(S_i) maps samples to matrices, not graphs to matrices, and the 'proof of bijection' assumes Φ is a function of S_i alone even though Definition 4.2 uses global sums over all samples and u-centered normalizations. Without a valid isomorphism, clustering in the kernel space has no proven causal semantics.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the nonlinear Causal Kernel Clustering (CKC) method for learning heterogeneous subgroups from observational data. The method constructs a u-centered sample mapping function Φ that compares a marginal distance covariance statistic with a chi-square threshold, builds a nonlinear causal kernel from these mappings, and clusters samples in the resulting kernel space. The central theoretical claim is that the sample mapping space is isomorphic to the causal graph space, which would guarantee that clusters correspond to differences in causal structure. The paper also reports experiments on synthetic data, Indian Ocean Dipole early warning, and Boston housing prediction, integrating the method with existing causal learning approaches.","tokens_in":17512,"tokens_out":6260,"duration_ms":62616,"significance":"If the causal identifiability claim were valid, the paper would offer a useful plug-and-play module for heterogeneous subgroup causal learning, with the attractive feature of using u-centered distance covariance to reduce finite-sample bias in nonlinear dependence measurement. The empirical work, especially the synthetic nonlinear experiments and the IOD case study, shows promise and is a real strength of the manuscript. However, the central theoretical guarantee is not established: the sign rule in Theorem 4.1 lacks a distributional basis, the proof of Theorem 5.1 reverses that sign rule and conflates graph connectivity with marginal dependence, and the bijectivity proof in Section 5.1 is unsupported. Because these issues are load-bearing for the paper's claim that clusters are causally meaningful, the theoretical contribution, which is the main novelty, does not hold as written.","major_comments":[{"comment":"The sign rule at the core of the method is asserted without a distributional basis. Φ(S_i) is defined as a sum of u-centered distance-covariance terms minus nχ²_{1−ν}(1), and Theorem 4.1 declares that a positive aggregate entry means dependence and a non-positive entry means independence. No argument shows that this difference (or its aggregate over samples) has the chi-square null distribution, that the chosen threshold controls a Type I error, or that the rule is consistent as n grows. Because Theorem 4.2 and Theorem 5.1 both inherit this sign rule, the causal interpretation of the clusters rests on an unvalidated thresholding procedure. A concrete fix would be to derive the asymptotic distribution of the aggregated statistic under H0, or to replace the threshold with a calibrated permutation test, and then state the resulting level and consistency properties.","section":"§4.2, Definition 4.2 and Theorem 4.1"},{"comment":"The proof of Theorem 5.1 reverses the sign convention established in Theorem 4.1. Theorem 4.1 says that a positive aggregate Φ entry implies X_p and X_q are not independent, but the proof of Theorem 5.1 states that a positive entry 'indicates a causal relationship between the features X_j and X_j′: X_j ⊥⊥ X_j′', i.e., independence. Independently, the proof equates m-connectivity in N_G with marginal dependence, which is false in general: in the collider X1→X3←X2, X1 and X2 have a path of length 2 and would be included in N_G for m=2, yet they are marginally independent and would receive a negative Φ entry. Hence sign(Φ) and the matrix I defined in Eq. (14) cannot agree in general, and the claimed isomorphism between causal graph space and causal matrix space is not established.","section":"§5.2, Theorem 5.1"},{"comment":"The bijectivity proof of Φ is not justified. It asserts that Φ(S_i)=Φ(S_i′) implies Z_i=Z_i′ and hence S_i=S_i′, but according to Definition 4.2 each Φ(S_i) depends on global quantities—sums over all α, β, ζ and dataset-wide normalizations—so Φ(S_i) is not a function of the single row S_i. The proof does not show injectivity, and the surjectivity claim ('for each η there exists S_ϵ with Φ(S_ϵ)=η') is asserted without construction. The isomorphism between the sample space and the mapping space, and with it the claim that clustering in the mapped space preserves sample-space structure, is therefore unproven.","section":"§5.1"},{"comment":"The claim that the method 'reduc[es] the predicted error in almost all scenarios' is contradicted by the reported RMSE values. For ERM, RMSE increases from 3.09 to 4.70, 4.69, and 4.68 for K=2, K=4, and K=5 respectively; for DWR, RMSE increases from 0.80 to 1.96 and 1.95 for K=2 and K=4. Only Sta_Error decreases in most scenarios. The table also does not display the underlines mentioned in the text. The evidence for improving downstream causal learning is thus only partial and should be reanalyzed with error bars or paired statistical tests.","section":"§6.2.2, Table 3"}],"minor_comments":[{"comment":"The notation is difficult to follow: Φ(S_i,) has a trailing comma, and V_{ζ,γ} uses Z_{ζ,γ,·} without a clear definition of the index ranges, especially since Z is introduced as an n×n×m tensor; please define all indices and dimensions explicitly.","section":"§4.2, Eqs. (9)–(11)"},{"comment":"The theorem states that Φ maps 'from the causal graph space Y_m to the causal matrix space G_m', but Definition 5.1 calls the causal graph space G_m and Definition 5.2 calls the causal matrix space Y_m; the names are reversed.","section":"Theorem 5.1 statement"},{"comment":"The header lists four condition columns under 'Linear Nonlinear', but each method row contains five V-measure/ARI pairs; the columns should be aligned and the linear versus nonlinear settings labeled clearly.","section":"Table 1"},{"comment":"The sentence 'providing 15 warnings out of 15 IOD events, with 13 of these predictions being accurate' is confusing given Table 2, which shows TP=13 and FN=2; please rephrase to clarify the relationship between warnings, events, and correct predictions.","section":"§6.2.1"},{"comment":"Reference [29] appears to have garbled author order ('Sambit Panda Cencheng Shen and Joshua T. V ogelstein'); the correct citation is likely C. Shen, S. Panda, and J. T. Vogelstein.","section":"References"}],"recommendation":"reject","confidential_remarks":"The empirical components, especially the synthetic nonlinear experiments and the IOD case study, show some promise, but the central theoretical claim is unsupported and the proof of Theorem 5.1 contains a sign reversal with respect to Theorem 4.1. Since the causal semantics of the clusters is the main advertised contribution, the load-bearing errors cannot be fixed by local revisions; a substantially reworked theoretical treatment and more careful empirical validation would be needed before reconsideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my read. The genuinely new part is the u-centered sample mapping function with the n-fold chi-square threshold, replacing the distance-covariance kernel from [10]. That is a plausible incremental step, and the synthetic nonlinear clustering results plus the IOD warning numbers give the method some empirical texture. The plug-and-play framing with stable-learning methods is also practically useful. That should be credited.\n\nBut the theory does not hold together. Theorem 5.1, the load-bearing claim that clusters in kernel space correspond to causal graph differences, is proven with a sign convention that contradicts Theorem 4.1. Theorem 4.1 says positive entries mean non-independence; Theorem 5.1's proof says positive entries mean a causal relationship and writes Xj ⊥⊥ Xj′ — the opposite. Then negative entries are said to mean no causal relationship while the text writes Xj ̸⊥⊥ Xj′. You cannot have both. Separately, m-connectivity records path existence up to length m, while Φ encodes marginal dependence. A collider X1→X3←X2 makes X1 and X2 marginally independent but m-connected for m≥2, so the indicator matrices I and sign(Φ) cannot agree in general. The isomorphism is also ill-typed: Φ maps samples to matrices, not graphs to matrices, and the bijection proof assumes Φ(S_i) depends only on S_i even though Definition 4.2 sums over all samples. These are not cosmetic; they are the entire causal identifiability argument.\n\nTheorem 4.1 lacks a distributional basis. Comparing a distance covariance statistic to a chi-square quantile without deriving a null distribution or consistency argument is not a valid test, so the binary dependence/independence rule that everything downstream inherits is unsupported.\n\nThe experiments are mixed. CKC beats baselines on nonlinear synthetic data, and the IOD result is interesting. But Table 3 undercuts the abstract: for ERM and DWR, RMSE gets worse after +CKC in every K; only Sta_Error improves. Claiming the method reduces prediction error is not supported by the paper's own numbers.\n\nWho should read it: someone working on heterogeneous causal clustering might mine the u-centered construction and the IOD setup, but they should treat the theory as open, not proven. My recommendation: this should not go to peer review in its current form. It needs either a correct isomorphism theorem or a repositioning as a heuristic method with no causal identifiability claims. If the authors do that, I would be happy to see it resubmitted.","headline":"Incremental kernel-clustering extension with a useful u-centered twist, but the central isomorphism proof reverses its own sign convention and the Boston experiments contradict the 'reduced prediction error' claim.","tokens_in":17944,"tokens_out":3401,"would_cite":false,"duration_ms":38387,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a u-centered sample mapping function turns each sample row into a matrix whose sign pattern encodes pairwise dependence and independence, and that this mapping space is isomorphic to the causal graph space, so…","keywords":["causal discovery","heterogeneous subgroups","causal clustering","kernel methods","distance covariance","u-centered estimator","causal identifiability","subgroup invariance"],"falsifier":"Generate many small samples of two independent standard normal features, compute the aggregated sample mapping entry with the paper's chi-square threshold at $\\nu = 0.05$, and check how often the sign is positive; if the rate is far above 5 percent, the sign rule is not controlling its stated error level and clusters formed by the sign pattern would not correspond to true dependence.","tokens_in":16915,"feed_emoji":"🔗","tokens_out":5931,"duration_ms":63270,"temperature":0.7,"pith_summary":"The paper is trying to establish that heterogeneous subgroups in observational data can be found by clustering samples in a space whose geometry mirrors the underlying causal graphs. It introduces a u-centered sample mapping function that turns each sample row into a matrix; a positive entry is read as dependence between the two features, a negative entry as independence. A nonlinear kernel on these matrices measures sample similarity, and a clustering step groups samples with matching causal sign patterns. The authors argue that because the mapping space is isomorphic to the causal graph space, the clusters are causally meaningful, not just statistically similar. They further show that feeding subgroup information into existing causal learning methods reduces prediction error on climate and housing data.","feed_headline":"Causal kernel clustering recovers heterogeneous subgroups from data","feed_subtitle":"A u-centered sample mapping makes clusters correspond to causal structure, improving downstream causal learning.","key_machinery":"The u-centered sample mapping function $\\Phi(S_i)$ together with the nonlinear causal kernel. The function subtracts an $n$-times chi-square threshold from a u-centered marginal distance covariance sum, and the sign pattern of the resulting matrix is the causal matrix. The kernel $\\kappa(S_i,S_{i'}) = \\langle \\Phi(S_i), \\Phi(S_{i'}) \\rangle_F / (\\|\\Phi(S_i)\\|_F \\|\\Phi(S_{i'})\\|_F)$ is a cosine similarity between the two mapping matrices, which turns sign-pattern agreement into cluster affinity. The load-bearing mechanism is the claimed isomorphism between the sign-pattern space and the causal graph space, which is what lets the authors interpret any clustering of the mapped samples as a grouping by causal structure.","core_discovery":"The central claim is Theorem 5.1: the sample mapping function $\\Phi(S_i)$ is an isomorphism between the causal graph space and the causal matrix space. Concretely, for each sample row $S_i$, $\\Phi(S_i)$ is an $m \\times m$ matrix whose $(p,q)$ entry is the difference between a marginal distance covariance term and an amplified chi-square threshold. The sign of that entry is the paper's binary decision: positive means $X_p$ and $X_q$ are nonlinearly dependent, negative means independent. Two samples are assigned to the same subgroup when their sign matrices match, and the paper argues this is equivalent to sharing the same m-connectivity pattern in the causal graph. This is what licenses the method's use of ordinary kernel clustering to recover causally heterogeneous subgroups.","pith_inferences":["Beyond the paper: the chi-square threshold could be calibrated against a permutation null on independent features, turning the sign rule into a formal hypothesis test with a controlled error rate.","Beyond the paper: if the isomorphism is taken seriously, a new environment's samples could be assigned to a subgroup purely by their sign matrix, making the method a causal-aware classifier for distribution shift.","Beyond the paper: the binary sign collapses continuous dependence strength, so a magnitude-weighted variant might separate subgroups that share an edge but differ in its strength, which the current method would merge.","Beyond the paper: the IOD early-warning result suggests using the sign of time-lagged causal kernels as a leading indicator for other climate indices, providing a transfer test of the underlying mechanism."],"forward_implications":["If the isomorphism holds, ordinary kernel clustering on the mapped samples returns clusters that reflect differences in causal relationships rather than differences in marginal distributions.","Because the method is a plug-and-play module, it can be prepended to existing causal structure learning and stable prediction methods; the paper demonstrates reduced prediction error on Boston housing for ERM, KerHRM, and stable learning baselines.","The sign of the aggregated time-lagged causal kernel between Indian Ocean Dipole regions can serve as an early warning signal about a year ahead, with the paper reporting 15 warnings, 13 correct out of 15 IOD events.","The method is robust to the choice of cluster number $K$, since what matters is capturing the heterogeneous subgroup information rather than the exact subgroup count.","No subgroup labels are needed in advance; the clustering discovers the causally distinct subgroups from the data."],"supporting_citations":[{"why":"Defines the distance covariance-based kernel for nonlinear causal clustering in heterogeneous populations, the method this paper extends with a u-centered estimator.","marker":"[10]"},{"why":"Introduces the u-centered distance matrix and partial distance covariance on which the unbiased sample mapping function is built.","marker":"[19]"},{"why":"Provides kernelized heterogeneous risk minimization and the causal clustering precedent used to justify integrating subgroup information into causal learning.","marker":"[12]"},{"why":"Supplies the Indian Ocean Dipole climate network data and early-warning baselines used in the real-world validation.","marker":"[47]"},{"why":"Cited for the chi-square independence test that sets the threshold in the sample mapping function's binary decision.","marker":"[28]"},{"why":"Cited for the chi-square test of distance correlation, supporting the n-amplified chi-square critical value in the mapping function.","marker":"[29]"},{"why":"Cited for adjacency-matrix representations of causal graphs, informing the causal matrix space construction.","marker":"[37]"}],"fun_headline_variants":["Causal kernel clustering maps data subgroups to causal structure","Sample mapping proves isomorphism for heterogeneous causal groups","Kernel clustering reveals causal subgroups via sign-matrix mapping","Nonlinear causal kernel clustering finds heterogeneous subgroups","Isomorphism-based clustering recovers heterogeneous causal subgroups"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that whenever its dependence measure exceeds a chi-square cutoff, the features are truly dependent, and whenever it falls below, they are truly independent, and no statistical guarantee for that comparison is given.","fun_headline_variants_meta":{"raw":{"variants":["Causal kernel clustering maps data subgroups to causal structure","Sample mapping proves isomorphism for heterogeneous causal groups","Kernel clustering reveals causal subgroups via sign-matrix mapping","Nonlinear causal kernel clustering finds heterogeneous subgroups","Isomorphism-based clustering recovers heterogeneous causal subgroups"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000687,"raw_usage":{"total_tokens":3061,"prompt_tokens":839,"completion_tokens":2222,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":2150}},"tokens_in":455,"tokens_out":2222,"duration_ms":16700,"temperature":1.0,"reasoning_tokens":2150,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:02:10.456132+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate many small samples of two independent standard normal features, compute the aggregated sample mapping entry with the paper's chi-square threshold at $\\nu = 0.05$, and check how often the sign is positive; if the rate is far above 5 percent, the sign rule is not controlling its stated error level and clusters formed by the sign pattern would not correspond to true dependence.","supporting_citations":[{"cited_title":"A Distance Covariance-based Kernel for Nonlinear Causal Clustering in Heterogeneous Populations","cited_arxiv_id":null,"evidence_quote":"Defines the distance covariance-based kernel for nonlinear causal clustering in heterogeneous populations, the method this paper extends with a u-centered estimator."},{"cited_title":"Székely and Maria L","cited_arxiv_id":null,"evidence_quote":"Introduces the u-centered distance matrix and partial distance covariance on which the unbiased sample mapping function is built."},{"cited_title":"Integrated Latent Heterogeneity and Invariance Learning in Kernel Space","cited_arxiv_id":null,"evidence_quote":"Provides kernelized heterogeneous risk minimization and the causal clustering precedent used to justify integrating subgroup information into causal learning."},{"cited_title":"Bogachev, and Juergen Kurths","cited_arxiv_id":null,"evidence_quote":"Supplies the Indian Ocean Dipole climate network data and early-warning baselines used in the real-world validation."},{"cited_title":"Seven proofs of the Pearson Chi-squared independence test and its graphical interpretation","cited_arxiv_id":"1808.09171","evidence_quote":"Cited for the chi-square independence test that sets the threshold in the sample mapping function's binary decision."},{"cited_title":"V ogelstein","cited_arxiv_id":null,"evidence_quote":"Cited for the chi-square test of distance correlation, supporting the n-amplified chi-square critical value in the mapping function."},{"cited_title":"High-dimensional joint estimation of multiple directed Gaussian graphical models","cited_arxiv_id":null,"evidence_quote":"Cited for adjacency-matrix representations of causal graphs, informing the causal matrix space construction."}],"review_version":1}