{"id":"42d3f203-6bf9-4e76-9491-8479b9f91980","arxiv_id":"2507.21792","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"MCVCI, a mixture conditional variational auto-encoder, is reported to identify causal direction in heterogeneous additive-noise data, and its residual-based clustering variant MCVCC recovers causal mechanism clusters with high ARI and NMI.","lead":"This paper proposes a neural network method, MCVCI, to decide which of two variables causes the other when the data mix several different causal mechanisms. It also introduces MCVCC, which groups data points by causal mechanism, and reports top accuracy on simulated and real benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Identifiability proof for HANM is invalid: the log of a mixture is replaced by a sum of logs and component-specific residuals are collapsed, so the likelihood-ratio criterion lacks theoretical support.","rationale":"The reader's weakest assumption correctly identifies the proof of Lemma 1/Theorem 1 as the load-bearing point. My independent reading confirms the proof gap is even more concrete than 'collapsing sums': the paper replaces the logarithm of a mixture by a weighted sum of log-densities, and then collapses component-specific residuals into a single residual. Both steps are mathematically unsupported. Since the ELBO-based decision criterion is presented as a direct consequence of this identifiability result, the central claim lacks a valid theoretical foundation. The secondary issue in Algorithm 1 (the corr>0 gate that skips likelihood computation for non-positive correlations) further weakens the practical claim, but the foundational proof gap is the primary concern. I therefore agree with the REJECT verdict and see no reason to change it.","tokens_in":13980,"tokens_out":7874,"duration_ms":97779,"concrete_test":"Use symbolic differentiation (e.g., SymPy) on a concrete K=2 Gaussian HANM with f_1(x)=x^2 and f_2(x)=sin(x). Compute the exact mixed partial derivative ∂²π/∂X∂Y and ∂²π/∂X² for the forward log-density log Σ_k w_k p_{ε_k}(y - f_k(x)), then evaluate ∂/∂X[ (∂²π/∂X²) / (∂²π/∂X∂Y) ]. If the result is not identically zero for generic w_k and noise variances, the collapse in Eqs. (10)-(12) is invalid and Lemma 1's derivation is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Lemma 1/Theorem 1, which asserts that a reverse HANM is 'almost impossible'. The proof in Section III-C is invalid at two algebraic steps. First, Eq. (8) writes the log-joint as π(X,Y) = Σ_k w_k v_k(y_k - f_k(x_k)) + ξ(X). But p(X,Y) is a mixture density, so the correct expression is ξ(X) + log Σ_k w_k p_{ε_k}(y_k - f_k(x_k)); the log of the sum is replaced by a weighted sum of logs, which is unjustified. Second, Eq. (10) collapses Σ_k w_k ev''_k(x_k - g_k(y_k)) g'_k(y_k) into a single function ev''(X - g(Y)), which can hold only if all g_k are identical. Because the ODE condition and the 'almost impossible' genericity claim derive entirely from these two steps, the identifiability theorem is unproven. The decision rule in Eqs. (25)-(26) is therefore not connected to any valid theoretical guarantee. If the flexible conditional VAE can approximate the reverse direction well, the two ELBO scores can be nearly equal, so the comparison loses discriminative power. This directly undermines the claimed reliability of MCVCI and MCVCC.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MCVCI, a method for bivariate causal direction identification on heterogeneous observational data, and MCVCC, an extension for clustering causal mechanisms. The model is a mixture of additive noise models (HANM); the authors claim an identifiability result (Lemma 1/Theorem 1) stating that a reverse HANM is 'almost impossible', and use it to justify a likelihood-ratio decision rule computed via a mixture conditional variational auto-encoder (CVAE). MCVCC clusters the residuals of the selected direction. Experiments on SIM/SIM-G/SIM-ln, CEP, and BAFU air data report accuracy/ARI/NMI improvements over several baselines, including ANM-MM, CANM, and classical methods. The central theoretical claim, however, rests on a proof that contains elementary algebraic errors in the log-density and derivative calculations, and the 'almost impossible' assertion lacks any measure-theoretic argument. The decision rule therefore lacks a valid identifiability foundation, and the empirical results, while reported as favorable, are not connected to a sound theoretical justification.","tokens_in":14298,"tokens_out":2849,"duration_ms":37136,"significance":"If the HANM identifiability theorem and the ELBO-based decision rule were correct, the paper would offer a practically useful extension of additive-noise causal inference to heterogeneous data, and the clustering application would be a modest novel contribution. The experimental coverage is broad, with several external benchmarks (SIM, SIM-G, SIM-ln, CEP) and a comparison against many recent methods, including ANM-MM and CANM. These strengths are real and deserve credit. However, the theoretical pillar—the identifiability of the HANM—is not established: the proof of Lemma 1 replaces the logarithm of a mixture by a weighted sum of logarithms, collapses sums of component derivatives into single functions without justification, and asserts genericity without a measure. Since the likelihood-ratio criterion depends entirely on this theorem, the central claim is unsupported. The paper does not provide machine-checked proofs or reproducible code, and the reported performance cannot compensate for the absence of a valid theoretical guarantee.","major_comments":[{"comment":"The proof begins by writing the log-joint in the reverse model as π(X,Y)=Σ_k w_k v_k(y_k−f_k(x_k)) + ξ(X). But p(X,Y) is a mixture density, so log p(X,Y)=ξ(X)+log Σ_k w_k p_{ε_k}(y_k−f_k(x_k)). The paper replaces the log of the sum with the weighted sum of logs, which is not an identity. Every subsequent derivative in Eqs. (9)–(11) is therefore taken from an expression that is not the actual log-density; the derived ODE (7) and the conclusion of Lemma 1 do not follow.","section":"Section III-C, Eq. (8)"},{"comment":"Equation (10) collapses Σ_k w_k ev''_k(x_k−g_k(y_k)) g'_k(y_k) into a single function ev''(X−g(Y)). This step is valid only if all component functions g_k are identical (or under other unstated conditions that force the sum to form a single function of (X,Y)). No such condition is given. The same issue affects Eq. (11), where Σ_k w_k ev''_k is replaced by a single ev''. The derivation of Eq. (12) and the subsequent cancellation steps depends on this unjustified collapse, so the proof of Lemma 1 is incomplete at a load-bearing point.","section":"Section III-C, Eq. (10)"},{"comment":"The theorem concludes that 'it is almost impossible to exist a hybrid ANM satisfying the condition from Y→X' and states that the set of solutions log p_X is contained in a three-dimensional affine space. No measure-theoretic argument is provided: there is no definition of the space of distributions or functions over which a measure is being taken, no statement of the genericity assumption, and no proof that the constraint set has measure zero. The 'almost impossible' claim is asserted rather than demonstrated. Since the likelihood-ratio decision rule in Eqs. (25)–(26) is justified solely by this genericity, the rule lacks a theoretical guarantee.","section":"Section III-C, Theorem 1"},{"comment":"The generative model is internally inconsistent. Equation (3) defines Y as a weighted sum of component functions, Y=Σ_k w_k(f_k(x_k)+ε_k), which is a single functional relationship. In contrast, Eq. (5) and the subsequent proof treat p(Y|X) as a mixture density Σ_k w_k p_{ε_k}(...), and Eq. (22) again claims log p(Y|X)=Σ_k w_k log p_{ε_k}(...). A weighted sum of functions is not the same as a mixture of conditional densities, and the logarithm of a mixture is not the mixture of logarithms. This inconsistency affects both the identifiability proof and the ELBO derivation: the variational model in Fig. 1 and Eq. (23) uses a mixture decomposition, while the causal criterion in Eqs. (25)–(26) is written as a sum of log-likelihoods. The relationship between the fitted CVAE and the stated HANM is therefore unclear, and the clustering objective in Eq. (27), which uses the residual from the selected direction, inherits this ambiguity.","section":"Section III-B and III-D, Eqs. (3), (5), (22)"},{"comment":"The ELBO in Eq. (24) is a lower bound for log pθ(Y|X) under a specific conditional generative model, but the paper does not show that this lower bound is a faithful proxy for the HANM marginal log-likelihood when comparing the two directions. In particular, the KL term and the encoder distribution qφ are optimized separately for the forward and reverse models, and the comparison in Algorithm 1 (step 7) uses the sum log p(X)+ELBO vs. log p(Y)+ELBO. Without an analysis of the approximation gap, two nearly equal ELBOs can arise even when the true likelihoods differ; conversely, a flexible VAE might approximate the reverse direction well, making the scores close and the decision unreliable. The empirical confidence analysis in Section IV-A4 measures only the decision rate of the top k%, not the validity of the underlying likelihood comparison.","section":"Section III-D, Eqs. (23)–(26)"}],"minor_comments":[{"comment":"The confidence paragraph contains a duplicated sentence: 'The higher the value of τ, the more correct our decision will be. Furthermore, we can set a threshold t to require τ ≥ t.' appears twice in a row. This should be cleaned up.","section":"Section IV-A4"},{"comment":"Several typos and nonstandard notations occur, e.g., 'casual' for 'causal' in the abstract and Algorithm 1 outputs, 'eϵ' vs. 'ε' inconsistency, and 'wϵc' in Section III-E. Please proofread carefully.","section":"Throughout"},{"comment":"The header of Table IV repeats 'DIFFERENT CLUSTER NUMBERS CONDITION' from Table III; the actual setting (BAFU air data) should be stated in the caption.","section":"Table IV"},{"comment":"The reference list is incomplete or inconsistent in places: reference [32] is formatted as 'S., Hoyer, et al.' with an unusual author order, and several cited works (e.g., the ANM-MM paper, the PNL paper) would benefit from full bibliographic data.","section":"References"},{"comment":"The description of the constructed simulated datasets is terse: it does not specify how the component index k is assigned to samples, how many samples per component are drawn, or how the mixture weights w_k are chosen. Please provide the full data-generation protocol so that the clustering experiments can be reproduced.","section":"Section IV-B2"}],"recommendation":"reject","confidential_remarks":"The paper has a promising empirical idea, but the theoretical foundation is unsound at an elementary level: the proof of Lemma 1 uses incorrect algebra for log-mixture densities, and the 'almost impossible' claim is not supported by any measure-theoretic argument. Since the central decision rule and the clustering method are both justified by this theorem, the contribution is not established. The errors are not fixable within a minor revision; they require a new proof or a different theoretical basis. The experimental section is extensive, and if the authors can provide a correct identifiability result or reframe the method as purely heuristic with appropriate caveats, a resubmission could be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the paper with the stress-test note in hand, and I think the stress-test hits the right nail. The core theoretical claim—that reverse HANM is 'almost impossible'—is not proven. In Eq. (8) the log of a mixture density is replaced by a sum of logs, which only holds if you assume a single component. The later collapse of Σ_k w_k ev''_k(...)g'_k(y_k) into a single function ev''(X - g(Y)) is also unjustified unless all g_k are identical. These are not minor gaps; the ODE condition and the genericity statement derive entirely from those steps. So the causal decision rule in Eqs. (25)-(26) has no valid identifiability guarantee to lean on.\n\nThat said, there is something here worth a referee's time. The idea of using a mixture conditional VAE to score bivariate causal directions, and then clustering on the residual of the chosen direction, is a genuine algorithmic combination not present in ANM-MM or CANM. The empirical results on SIM, SIM-G, SIM-ln, and CEP are strong if they hold up, and the authors compare against a wide range of baselines. The clustering experiments, while noisy in reporting, suggest the residual-based features carry real signal.\n\nThe soft spots beyond the proof: the link between the ELBO criterion and the identifiability theorem is never made explicit—ELBO is a lower bound, and nothing shows that comparing bounds in two directions equals comparing true likelihoods. With a flexible VAE, both directions can fit the same joint distribution, which can erode the difference. Also, no code is released, and the experimental tables have inconsistencies (e.g., ARI/NMI values that look like percentages but with odd formatting, and some suspicious perfect scores). These don't automatically sink the empirical claim but they make it hard to trust without reimplementation.\n\nWho is this for? A researcher working on bivariate causal discovery with heterogeneous data would want to know about this combination, and the empirical benchmark comparison is useful. But they should treat the theoretical guarantee as unproven until fixed.\n\nRecommendation: send it to peer review. The algorithmic idea and empirical scope deserve a serious look, but it needs major revision: either a corrected identifiability proof that treats the mixture correctly, or a removal of the theoretical claim and a framing of the method as a heuristic with strong empirical performance. A referee should be asked to focus on the proof and on the reproducibility of the tables.","headline":"The algorithmic combination is interesting and the empirical results are strong, but the identifiability proof has a fatal algebraic gap, so the causal decision rule currently rests on sand.","tokens_in":14781,"tokens_out":2716,"would_cite":false,"duration_ms":32094,"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":"A mixture of additive noise models is identifiable from pooled observational data, so comparing variational likelihood bounds in the two directions finds the true causal direction, and clustering the residual reveals the generating…","keywords":["hybrid additive noise model","bivariate causal discovery","mixture conditional variational autoencoder","ELBO likelihood criterion","causal mechanism clustering","heterogeneous observational data","identifiability","causal inference"],"falsifier":"Generate data from a known forward HANM with two near-linear components and small noise, then run MCVCI many times; if the forward and reverse ELBO scores are statistically indistinguishable across seeds, the reverse-identifiability premise fails.","tokens_in":13775,"feed_emoji":"🔀","tokens_out":8444,"duration_ms":91893,"temperature":0.7,"pith_summary":"The paper claims that bivariate causal direction remains identifiable when the data come from a mixture of additive noise mechanisms, not just a single mechanism. It constructs a mixture conditional variational auto-encoder whose evidence lower bound approximates the conditional likelihood in each direction, and decides X→Y versus Y→X by whichever direction has the larger bound. In the chosen direction, the residual noise term carries the identity of the generating mechanism, so clustering that residual separates samples by cause rather than by geometric proximity. If these claims hold, heterogeneous observational data pooled across environments or locations—without environment labels—can be used for causal discovery and for clustering data by causal mechanism.","feed_headline":"Mixture noise models reveal cause direction and cluster mechanisms","feed_subtitle":"A dual-direction likelihood gap in a mixture conditional VAE finds the true cause and groups samples by mechanism.","key_machinery":"The load-bearing object is the Hybrid Additive Noise Model (HANM), a weighted sum of $K$ additive noise mechanisms, together with the mixture conditional variational auto-encoder built to fit it. The identifiability argument reduces the existence of a reverse HANM to a third-order ordinary differential equation (Eq. 7) on $\\xi = \\log p(X)$: generically no reverse model can satisfy it, so the likelihood gap between the ELBO bounds (Eqs. 25 and 26) is the decision rule. For clustering, the residual $w\\epsilon$ of the chosen direction is the extracted causal feature space, and the clustering objective minimizes the squared distance from residuals to cluster centers in that space.","core_discovery":"The central discovery is a generic identifiability result for the Hybrid Additive Noise Model (HANM), written as $Y = \\sum_{k=1}^K w_k(f_k(x_k) + \\epsilon_k)$ with $\\sum w_k = 1$ and $x_k \\perp\\!\\!\\perp \\epsilon_k$. The paper proves that if a forward HANM holds, a reverse HANM $X = \\sum_{k=1}^K w_k(g_k(y_k) + \\tilde\\epsilon_k)$ can exist only when the log-density of $X$ satisfies a specific third-order ordinary differential equation constraint; generically no such reverse model exists. Building on this, the mixture conditional variational auto-encoder maximizes an ELBO that lower-bounds $\\log p(Y|X)$, and the same bound in the reverse direction lower-bounds $\\log p(X|Y)$; comparing the two scores gives the causal direction. The residual $w\\epsilon$ in the winning direction is then used as the feature space for causal mechanism clustering.","pith_inferences":["A direct stress test the paper does not run: generate near-linear, low-noise two-component mixtures where the reverse regression is almost as simple as the forward one, and check whether the two ELBO scores become statistically indistinguishable; if they do, the generic identifiability claim needs a measure-theoretic qualification.","The residual-clustering idea suggests a practical extension: the same trained model could detect unknown environment switches or label shifts by inspecting whether residuals naturally split into more clusters than the assumed $K$.","Using the confidence threshold $\\tau$ as an abstention rule in downstream applications would convert the top-k decision-rate curve into a deployable policy, but the paper does not propose such a procedure."],"forward_implications":["Bivariate causality on pooled heterogeneous data can be decided without knowing which environment each sample came from.","Cluster assignments can be derived from the data-generating mechanism rather than from raw feature distance, separating shifts in function or noise.","The likelihood gap defines a confidence score $\\tau$, so users can set a threshold and abstain on low-confidence pairs instead of committing to a direction.","The variational regressor does not require Gaussian noise within a component, so the method covers nonlinear functions and non-Gaussian additive noise.","On the SIM, SIM-G, SIM-ln, and cause-effect-pair benchmarks the reported direction accuracy exceeds the comparison methods, and on the simulated and air-quality clustering tasks the reported ARI and NMI exceed the baselines."],"supporting_citations":[{"why":"Supplies the additive noise model whose asymmetry HANM extends.","marker":"[6]"},{"why":"Provides the mixture-of-ANMs identifiability proof template and the causal mechanism clustering baseline.","marker":"[12]"},{"why":"Supplies the conditional variational auto-encoder architecture and ELBO formulation that the likelihood criterion is built on.","marker":"[15]"},{"why":"Contributes the variational log-likelihood approach to causal direction that MCVCI adapts to the mixture setting.","marker":"[23]"},{"why":"Provides the simulated and real cause-effect benchmark datasets used for evaluation.","marker":"[24]"}],"fun_headline_variants":["Mixture conditional VAE identifies causal direction","Causal direction from mixture noise identifiability","Heterogeneous causality via mixture VAE bounds","Cluster causal mechanisms with mixture VAE","Hybrid noise model reveals cause and clusters"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof that a reverse hybrid additive noise model is almost impossible collapses the mixture's $K$ components into single functions in Eqs. (10)-(11) and asserts genericity without a formal measure argument, so if the flexible neural regressor can approximate the reverse direction anyway, the two likelihood scores become nearly equal and the direction decision loses its basis.","fun_headline_variants_meta":{"raw":{"variants":["Mixture conditional VAE identifies causal direction","Causal direction from mixture noise identifiability","Heterogeneous causality via mixture VAE bounds","Cluster causal mechanisms with mixture VAE","Hybrid noise model reveals cause and clusters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000639,"raw_usage":{"total_tokens":2938,"prompt_tokens":932,"completion_tokens":2006,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":1939}},"tokens_in":548,"tokens_out":2006,"duration_ms":16825,"temperature":1.0,"reasoning_tokens":1939,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:22:47.084803+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate data from a known forward HANM with two near-linear components and small noise, then run MCVCI many times; if the forward and reverse ELBO scores are statistically indistinguishable across seeds, the reverse-identifiability premise fails.","supporting_citations":[{"cited_title":"Nonlinear causal discovery with additive noise models,","cited_arxiv_id":null,"evidence_quote":"Supplies the additive noise model whose asymmetry HANM extends."},{"cited_title":"Causal inference and mechanism clustering of a mixture of additive noise models,","cited_arxiv_id":null,"evidence_quote":"Provides the mixture-of-ANMs identifiability proof template and the causal mechanism clustering baseline."},{"cited_title":"Learning structured output representation using deep conditional generative models,","cited_arxiv_id":null,"evidence_quote":"Supplies the conditional variational auto-encoder architecture and ELBO formulation that the likelihood criterion is built on."},{"cited_title":"Causal Discovery with Cascade Nonlinear Additive Noise Models","cited_arxiv_id":"1905.09442","evidence_quote":"Contributes the variational log-likelihood approach to causal direction that MCVCI adapts to the mixture setting."},{"cited_title":"Distinguishing cause from effect using obser- vational data: methods and benchmarks,","cited_arxiv_id":null,"evidence_quote":"Provides the simulated and real cause-effect benchmark datasets used for evaluation."}],"review_version":1}