{"id":"06c50fe0-7aa9-4f3f-bdb3-6e659bf7bb09","arxiv_id":"2502.08549","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An iterative algorithm (CBMM-GICE) automatically selects marginal and copula distribution forms in mixture models and improves clustering over Gaussian mixture models on MNIST and cardiac MRI data.","lead":"This paper adapts an existing iterative estimation algorithm, GICE, to automatically choose the mixture component distributions, including non-Gaussian shapes, for clustering. It tests the method on synthetic data, the MNIST database, and cardiac MRI scans, reporting better accuracy and fit than standard Gaussian mixture models in most comparisons.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No convergence or stability guarantee for GICE is established; with stochastic label simulation, fixed iterMax=100, and data-dependent UMAP, the claimed general superiority over GMM-EM is not yet supported.","rationale":"I agree with the reader that the absence of a convergence guarantee for GICE is the weakest assumption. The paper explicitly acknowledges this in Section 3.2 and Section 5. The algorithm is stochastic (T simulated labelings), uses discrete form selection, and a fixed iteration budget; without a proof or a careful stability analysis, the claimed general superiority over GMM-EM could be an artifact of the chosen seeds, initialization, or UMAP embedding. The empirical evidence in Table 3 has overlapping ranges, and the synthetic experiments lack a direct GMM-EM comparison. I also note the swapped assignments in Algorithm 1 lines 7-8, which should be corrected and verified, but this is likely a typographical issue. The appropriate verdict remains CONDITIONAL: the paper should be accepted only after the authors release code, fix the pseudocode, and provide a seed-sensitivity analysis with confidence intervals for the key comparisons. My concern does not overturn the positive direction of the results, so I recommend no change to the reader's conditional verdict.","tokens_in":18737,"tokens_out":9998,"duration_ms":109716,"concrete_test":"Run a stability audit: on the MNIST UMAP projection and the cardiac LAD data, repeat CBMM-GICE and GMM-EM with 50 independent random seeds (same UMAP projection) and, separately, 50 fresh UMAP projections, recording accuracy, Kolmogorov distance, and selected forms per run. Compute the paired difference CBMM-GICE minus GMM-EM with a 95% bootstrap confidence interval. If the interval includes zero, or if the selected marginal/copula forms vary substantially across seeds, the general outperformance and 'appropriate CBMM' claims are not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two parts: that GICE identifies an appropriate CBMM for any cohort, and that CBMM-GICE generally outperforms GMM-EM. Both rest on the iterations of Algorithm 1 settling at a useful model. Section 3.2 states \"It is difficult to study theoretically the convergence of GICE\", and Section 5 reiterates that convergence is only demonstrated indirectly. The algorithm simulates T label realizations per iteration (T=10 on MNIST, T=50 on cardiac data), performs discrete form selection by Kolmogorov distance, and is stopped at iterMax=100. Because there is no monotone objective (unlike EM likelihood), the sequence can in principle cycle or depend on the random seed and initialization. The reported empirical support is limited: Table 3 shows overlapping min-max accuracy ranges between GMM-EM and CBMM-GICE across 20 runs, and the synthetic experiments in Section 4.1 compare GICE variants but do not include a GMM-EM baseline, despite the conclusion claiming outperformance was tested on synthetic data. An internal inconsistency also appears in Algorithm 1, where lines 7-8 swap the marginal and copula estimation calls relative to Algorithms 2/3 and the text, making the exact procedure ambiguous without code. Together, these gaps mean the central claim is not yet convincingly established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CBMM-GICE, an adaptation of the Generalized Iterative Conditional Estimation (GICE) algorithm for unsupervised identification of copula-based mixture models (CBMMs) with heterogeneous marginal and copula forms. The algorithm iteratively simulates latent labels, selects marginal and copula forms from finite dictionaries, estimates parameters, and monitors convergence through a Kolmogorov distance index. The method is evaluated on synthetic two-cluster data, the full MNIST database after UMAP projection, and cardiac MRI infarct-pattern data, where it is compared with GMM-EM and, in one experiment, MMST-EM. The authors conclude that CBMM-GICE generally outperforms classic EM-identified mixtures and can identify subgroups with non-Gaussian shapes.","tokens_in":19061,"tokens_out":5666,"duration_ms":59048,"significance":"If the empirical claims are supported, the contribution is practically valuable: it provides a flexible parametric clustering method that performs marginal and copula form selection inside an iterative estimation scheme, avoiding exhaustive model-combination searches. The paper tests the method on a large public dataset and a clinical cohort, and its discussion of limitations is candid. The current evidence, however, does not yet fully support the breadth of the claims: there is no convergence guarantee, the pseudocode is internally inconsistent at one point, and several experimental comparisons lack uncertainty quantification and relevant baselines. The contribution is therefore promising but conditional on addressing these issues.","major_comments":[{"comment":"The central claim that CBMM-GICE reliably identifies a useful CBMM presupposes that the stochastic iterations settle at a stable model. The paper states in Section 3.2 that 'It is difficult to study theoretically the convergence of GICE' and only monitors convergence via the Kolmogorov distance in Eq. (15), with a fixed iterMax=100 and stochastic label simulation with T=10 or T=50. Because no monotone objective is optimized, the sequence could in principle cycle or depend on the random seed, UMAP projection, and initialization. Table 3 reports min-max accuracy ranges that overlap substantially between GMM-EM (0.690-0.956) and CBMM-GICE (0.716-0.967), and Table 4 reports single values without repeated runs. Please provide stability evidence: repeated seeds and initializations with full distributions, sensitivity to T and iterMax, and a check for fixed points or cycles.","section":"Section 3.2 and Algorithm 1"},{"comment":"The pseudocode in Algorithm 1 assigns alpha_k = MarginEstimation and theta_k = CopulaEstimation, which is the reverse of the definitions in Section 3 (steps 2(c)-2(d)) and of Algorithms 2 and 3, where MarginEstimation returns theta_k and CopulaEstimation returns alpha_k. This ambiguity affects the exact procedure being evaluated and must be resolved. Please correct the pseudocode or the text and state explicitly which quantity is selected by the decision rules Delta_1 and Delta_2 at each step.","section":"Algorithm 1, lines 7-8"},{"comment":"The conclusion states that CBMM-GICE was tested against GMM-EM 'both on synthetic and real medical imaging data', but the synthetic experiments in Section 4.1 compare only GICE variants with different T and initialization and include no GMM-EM baseline. The claimed general outperformance over GMM-EM is therefore not directly supported on synthetic data. Please add a GMM-EM comparison to the synthetic experiments or temper the conclusion accordingly.","section":"Section 4.1 and Section 5"},{"comment":"The cardiac evaluation is entirely in-sample: the UMAP projection is selected by trustworthiness on the same cohort, K=3 is selected by BIC on the same LAD data, and both Kolmogorov distance and silhouette scores are computed on the data used for fitting. Table 4 reports a single run without error bars or repeated UMAP projections/seeds. The reported improvements (silhouette 0.513 vs 0.452; Kolmogorov distance 0.059 vs 0.066/0.067) therefore lack uncertainty quantification and may reflect in-sample fit. Please add repeated projections and seeds and report confidence intervals or significance tests.","section":"Sections 4.3.1-4.3.2, Table 4 and Figure 5"},{"comment":"The paper identifies Online EMDL and VCMM as closely related CBMM identification methods but never compares with them in the experiments. Since these methods directly address the same marginal and copula form-selection problem, a comparison on at least the synthetic data, or an explicit statement of unavailability, would substantially strengthen the claim that the proposed identification strategy is original and advantageous.","section":"Section 1 and references [13,37]"}],"minor_comments":[{"comment":"The header 'GMM-EN' should read 'GMM-EM'.","section":"Table 3"},{"comment":"The text reports running time on a '2.7MHz CPU'; this should presumably be '2.7 GHz CPU'.","section":"Section 4.1.1"},{"comment":"The caption says 'clusters found be GMM-EM and CBMM-GICE'; it should read 'found by GMM-EM and CBMM-GICE'.","section":"Figure 5 caption"},{"comment":"The parameter range for Arch12 and Arch14 is written as alpha in [1,+infinity]; if the endpoint is not attained, use [1,infinity) for clarity.","section":"Table A2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the journal's scope and contains a plausible algorithmic contribution. The main obstacles are the Algorithm 1 inconsistency, the lack of uncertainty quantification in several experiments, and the missing comparison with previously proposed CBMM identification methods. I do not see grounds for rejection, as the method is clearly described apart from the pseudocode swap and the authors acknowledge several limitations. A full convergence proof is not necessary, but empirical stability evidence should be provided."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a credible methods paper, not a breakthrough. The new piece is the GICE adaptation to CBMM identification with Kolmogorov-distance form selection, plus the first real study of the realization time T. The synthetic experiments do show GICE recovering non-Gaussian marginals and copulas reasonably well, and the MNIST run at N=70000 is a serious stress test. The cardiac application is sensible and the authors are honest about its limits. I would not desk-reject this.\n\nThe soft spots are real but proportionate. The paper openly says convergence of GICE is theoretically intractable, and the only monitor is an in-sample Kolmogorov distance with a fixed iterMax=100. Given T label simulations and form selection at each iteration, the algorithm could cycle or be seed-dependent; nothing in the paper rules that out. The empirical support is also thinner than the conclusion claims. The synthetic section compares GICE variants only—no GMM-EM baseline appears there, even though the conclusion says CBMM-GICE generally outperforms GMM-EM on synthetic and real data. On MNIST, Table 3 shows overlapping min–max accuracy ranges, so the advantage is directional, not decisive. The cardiac results have no error bars, and the UMAP projection is selected with a data-dependent trustworthiness criterion, which makes the comparison to GMM-EM less clean.\n\nThere is also an internal inconsistency worth flagging: Algorithm 1 lines 7–8 swap the marginal and copula estimation calls relative to Algorithm 2, Algorithm 3, and the surrounding text. The exact procedure is ambiguous without code, and code is only promised upon acceptance. That ambiguity matters for a methods paper.\n\nI disagree with the strongest form of the stress-test note: the paper does not rest on a claim that GICE always converges to the true model. The authors explicitly say it does not ensure finding the true forms and discuss local extrema. But the weaker claim—that CBMM-GICE generally outperforms GMM-EM—is not yet convincingly established at the stated strength.\n\nWho is this for? Researchers working with copula mixtures in imaging who want a flexible, interpretable alternative to GMM. They will get a useful, reasonably honest empirical study. It should go to peer review, but with three demands: release code, compare against existing CBMM identification methods such as Online EMDL and VCMM, and fix Algorithm 1. With those changes, the paper could be a solid contribution to a subfield.","headline":"A plausible, clearly written adaptation of GICE to copula-based mixture identification with useful experiments, but the headline claim overreaches: convergence is unproven, a key baseline is missing from the synthetic section, and an ambiguity in Algorithm 1 needs fixing before it can be trusted.","tokens_in":19542,"tokens_out":1427,"would_cite":false,"duration_ms":17764,"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":"An unsupervised algorithm, CBMM-GICE, identifies copula-based mixture models with cluster-specific marginals and copulas and generally outperforms Gaussian-mixture EM on synthetic, MNIST, and cardiac MRI data.","keywords":["copula-based mixture models","GICE algorithm","model selection","finite mixtures","clustering","medical imaging","unsupervised learning","MNIST"],"falsifier":"Run CBMM-GICE on the synthetic two-cluster CBMM ground truth of Section 4.1 (N=2000, known parameters) from many random seeds, with both K-Means and GMM initialization and T=1 versus T=10, and compare the selected marginal and copula forms and label error ratios at iteration 100; if the final forms or error ratios vary materially across seeds on the same fixed data, the method's reported advantage over GMM-EM would not transfer reliably to new cohorts.","tokens_in":18541,"feed_emoji":"🧩","tokens_out":11721,"duration_ms":101792,"temperature":0.7,"pith_summary":"This paper proposes an unsupervised algorithm, CBMM-GICE, that builds a copula-based mixture model for a cohort of samples and clusters the samples into subgroups at the same time. Unlike a Gaussian mixture model, each cluster is allowed to have its own marginal distribution per dimension and its own dependence structure (copula), so skewed, heavy-tailed, or otherwise non-elliptical subgroups can be represented without artificially inflating the number of components. The authors argue that this flexibility matters in medical imaging, where patient subgroups often do not follow the same distribution family. They report that CBMM-GICE generally outperforms GMM-EM on synthetic two-cluster data, on the full MNIST database, and on cardiac magnetic resonance data for infarct territory and pattern clustering.","feed_headline":"Copula mixture model beats GMM on MNIST and cardiac MRI","feed_subtitle":"Each cluster gets its own marginal and copula forms, lifting MNIST accuracy to 0.848 from 0.824.","key_machinery":"The central object is the copula-based mixture model built on Sklar's theorem, which writes each component density as a copula density acting on the marginal CDFs times the product of the marginal densities: $p_k(x) = c_k(F_{k,1}(x_1),\\dots,F_{k,D}(x_D)) \\prod_{d=1}^{D} f_{k,d}(x_d)$. The machinery that carries the argument is the GICE loop: simulate the hidden labels $T$ times from the current posteriors, fuse the $T$ subgroups for each component, then apply Kolmogorov-distance decision rules to pick one marginal form per dimension from a dictionary $\\mathcal{H}$ and one copula form from a dictionary $\\mathcal{G}$, estimating parameters with maximum likelihood for marginals and pseudo-maximum likelihood for copulas. This form-selection step replaces the combinatorial search over all combinations of candidate forms. The realization time $T$ smooths the stochastic convergence at the price of runtime.","core_discovery":"The paper's central claim is that the Generalized Iterative Conditional Estimation (GICE) algorithm, originally designed for switching hidden Markov models, can be adapted to identify Copula-Based Mixture Models (CBMMs) in an unsupervised way. At each iteration the algorithm simulates hidden cluster labels from the current model, groups the samples accordingly, and then selects — for each cluster, each dimension, and the cluster's dependency structure — the best marginal and copula forms from candidate dictionaries, re-estimating their parameters as it goes. Because the candidate list includes Gaussian marginals and the Gaussian copula, GMMs are a special case of the search space. The authors report that on the full MNIST database CBMM-GICE reaches an average clustering accuracy of 0.848 versus 0.824 for GMM-EM, and on cardiac infarct data it produces lower Kolmogorov distances and higher silhouette scores, supporting the conclusion that heterogeneous, non-elliptical components fit real cohorts better.","pith_inferences":["Editorial inference: because convergence is only monitored and not proven, a natural stress test is to run many random seeds on the Section 4.1 synthetic ground truth and measure how often the selected marginal/copula forms and the label error ratio settle on the same answer.","Editorial inference: the real-data experiments all work in a 2D UMAP projection, so the reported accuracy reflects the projection quality as well as the mixture identification; comparing GMM-EM on the same coordinates, as the paper does, is the right way to isolate the mixture-model contribution.","Editorial inference: if the form-selection loop proves stable, the same scheme could extend to higher-dimensional latent spaces using vine or neural copulas, the direction the authors flag as the main obstacle to going beyond 2D."],"forward_implications":["Since Gaussian marginals and Gaussian copulas are inside the candidate dictionaries, GMM is a special case of the model family CBMM-GICE searches, so the method can only gain flexibility relative to GMM-EM on the same data.","On the full MNIST database the method reports an average clustering accuracy of 0.848 against 0.824 for GMM-EM, with a best run of 0.967, indicating that the added distributional flexibility helps on overlapping clusters.","On cardiac infarct data the method reports lower Kolmogorov distances than GMM-EM and MMST-EM and the highest mean silhouette score (0.513 versus 0.452), which matters for subgroup discovery in clinical cohorts.","The realization time T provides a practical trade-off: larger T gives smoother convergence and stabler estimates, but runtime grows from about 207 s at T=1 to 1548 s at T=10 in the synthetic test."],"supporting_citations":[{"why":"Supplies the iterative conditional estimation framework that CBMM-GICE adapts, including the two assumptions about complete-data estimators and simulation of hidden labels.","marker":"[33]"},{"why":"Introduces the Generalized ICE algorithm for switching hidden Markov models, the direct predecessor whose realization-time parameter T the paper studies.","marker":"[42]"},{"why":"Extends GICE to hidden Markov chains with unknown noise copulas and margins, providing the pseudo-maximum-likelihood copula estimation and form-decision precedent.","marker":"[8]"},{"why":"Provides the copula theory underlying the decomposition of joint distributions into marginals and dependence structures.","marker":"[32]"},{"why":"States Sklar's theorem, the identity that lets every multivariate distribution be written as a copula linking its marginals.","marker":"[39]"},{"why":"Presents Online EMDL, the closest prior method for online marginal and copula selection, which the paper contrasts with its batch form-updating GICE.","marker":"[13]"},{"why":"Defines the EM algorithm used to produce the GMM baseline that CBMM-GICE is compared against.","marker":"[30]"},{"why":"Supplies the full MNIST database used for the 10-cluster real-data comparison.","marker":"[22]"},{"why":"Supplies the UMAP projection used to bring MNIST and cardiac imaging data into two dimensions before clustering.","marker":"[29]"},{"why":"Provides the myocardial infarct pattern data and the pixel-wise analysis pipeline used in the cardiac imaging application.","marker":"[10]"}],"fun_headline_variants":["Copula clustering outshines GMM on MNIST and cardiac MRI","CBMM-GICE: heterogeneous copulas win on imaging data","Flexible copulas boost clustering accuracy on MNIST","GICE finds copula mixtures that beat GMM on MRI","Copula-based mixtures: better clustering from imaging data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the algorithm's random label-simulation loop settles on a stable, useful mixture model in practice; the paper gives no theoretical convergence proof and only monitors a goodness-of-fit index over at most 100 iterations, so different random seeds or starting points could in principle lead to different models.","fun_headline_variants_meta":{"raw":{"variants":["Copula clustering outshines GMM on MNIST and cardiac MRI","CBMM-GICE: heterogeneous copulas win on imaging data","Flexible copulas boost clustering accuracy on MNIST","GICE finds copula mixtures that beat GMM on MRI","Copula-based mixtures: better clustering from imaging data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000536,"raw_usage":{"total_tokens":2574,"prompt_tokens":942,"completion_tokens":1632,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":1548}},"tokens_in":558,"tokens_out":1632,"duration_ms":10577,"temperature":1.0,"reasoning_tokens":1548,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T04:40:32.206719+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run CBMM-GICE on the synthetic two-cluster CBMM ground truth of Section 4.1 (N=2000, known parameters) from many random seeds, with both K-Means and GMM initialization and T=1 versus T=10, and compare the selected marginal and copula forms and label error ratios at iteration 100; if the final forms or error ratios vary materially across seeds on the same fixed data, the method's reported advantage over GMM-EM would not transfer reliably to new cohorts.","supporting_citations":[{"cited_title":"Pieczynski, Convergence of the iterative conditional estimation and ap- plication to mixture proportion identification, Proc","cited_arxiv_id":null,"evidence_quote":"Supplies the iterative conditional estimation framework that CBMM-GICE adapts, including the two assumptions about complete-data estimators and simulation of hidden labels."},{"cited_title":"Zheng, S","cited_arxiv_id":null,"evidence_quote":"Introduces the Generalized ICE algorithm for switching hidden Markov models, the direct predecessor whose realization-time parameter T the paper studies."},{"cited_title":"Derrode, W","cited_arxiv_id":null,"evidence_quote":"Extends GICE to hidden Markov chains with unknown noise copulas and margins, providing the pseudo-maximum-likelihood copula estimation and form-decision precedent."},{"cited_title":"Nelsen, An introduction to copulas, Springer Science & Business Media, 2007","cited_arxiv_id":null,"evidence_quote":"Provides the copula theory underlying the decomposition of joint distributions into marginals and dependence structures."},{"cited_title":"Sklar, Fonctions de r´ epartition ` a n dimensions et leurs marges, Publi- cations de l’Institut Statistique de l’Universit´ e de Paris 8 (1959) 229–231","cited_arxiv_id":null,"evidence_quote":"States Sklar's theorem, the identity that lets every multivariate distribution be written as a copula linking its marginals."},{"cited_title":"Fujimaki, Y","cited_arxiv_id":null,"evidence_quote":"Presents Online EMDL, the closest prior method for online marginal and copula selection, which the paper contrasts with its batch form-updating GICE."},{"cited_title":"McLachlan, T","cited_arxiv_id":null,"evidence_quote":"Defines the EM algorithm used to produce the GMM baseline that CBMM-GICE is compared against."},{"cited_title":"Lecun, L","cited_arxiv_id":null,"evidence_quote":"Supplies the full MNIST database used for the 10-cluster real-data comparison."},{"cited_title":"McInnes, J","cited_arxiv_id":null,"evidence_quote":"Supplies the UMAP projection used to bring MNIST and cardiac imaging data into two dimensions before clustering."},{"cited_title":"Duchateau, M","cited_arxiv_id":null,"evidence_quote":"Provides the myocardial infarct pattern data and the pixel-wise analysis pipeline used in the cardiac imaging application."}],"review_version":1}