{"id":"48ecc1e7-42eb-4a0b-8c50-cbebd87637bc","arxiv_id":"2509.03652","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper derives a rank-selection rule for nonnegative matrix factorization from a common-cause predictability inequality and finds that the resulting features are stable across noise and random seeds on several image datasets.","lead":"A machine learning paper connects nonnegative matrix factorization to an old principle from probabilistic causality, and uses that connection to choose the number of features in image data. The authors report that their rank choice is stable under noise and yields reproducible image features, which could make NMF more reliable in practice.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stability evidence for the central claim lacks a chance baseline and a semantic check for outlier pairs; the low mean matched distance in Section VI is insufficient to support 'effectively resolving nonidentifiability'.","rationale":"The reader's weakest assumption focuses on the representativeness of local minima computed for one seed pair and the lack of chance-level baselines. I partially agree, but I sharpen this to a concrete, testable deficiency: the stability evidence includes a long tail of very distant matched pairs, and the authors explicitly concede in Section IX that they cannot judge semantic equivalence. This is the most load-bearing point because the central claim is specifically about resolving nonidentifiability—if some basis images change completely across seeds, the resolution is incomplete even if the mean distance is low. The proposed test directly addresses both the missing null model and the outlier issue. A ground-truth validation on Swimmer (where true components are known) would also be valuable, but the permutation-null test is the minimal check that would settle whether the observed stability is meaningful. I therefore recommend keeping the reader's CONDITIONAL verdict: the contribution is promising and the heuristic is interesting, but the central claim requires stronger evidence before it can be accepted as stated.","tokens_in":18206,"tokens_out":12031,"duration_ms":125802,"concrete_test":"Run a permutation-null stability test on UTK at R=36: (i) obtain basis-image sets from 20 NMF runs with different seeds (same data); (ii) for each pair, compute Hungarian-matched cosine distances as in (26); (iii) construct a null distribution by applying the same matching to basis sets from NMF runs on two copies of UTK whose pixel indices have been independently permuted (preserving pixel marginals, destroying shared facial structure); (iv) compare observed mean/median/max to the null via a permutation test. If the observed distribution is not significantly below the null (e.g., p>0.05) or if more than 5% of matched pairs exceed distance 0.5, the stability claim—and hence the resolution of nonidentifiability—is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Abstract; Section VI) is that NMF around the R_c rank yields basis images stable against noise and seed, thereby effectively resolving nonidentifiability. The supporting evidence is limited: one train/test split of UTK, one seed pair per condition, and Hungarian-matched cosine distances without any null model. Table V shows that at R=36 (the rank closest to R_c), the seed-only experiment has max distance 0.9901 and min 0.0012, i.e., a long tail of almost-orthogonal matched pairs; the authors themselves state in Section IX that they lack semantic analysis tools to determine whether such numerically distant basis images are semantically equivalent. This matters because the claim is about resolving nonidentifiability: if a non-negligible fraction of basis images change completely across local minima, the solution is not effectively identified. Moreover, without a chance baseline, the low mean/median could be an artifact: basis images from the same face dataset may be structurally similar even under random local minima, and Hungarian matching will always produce some low distances. Thus the observed stability does not distinguish genuine identification from trivial similarity.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a conceptual and methodological connection between nonnegative matrix factorization (NMF) and Reichenbach's principle of the common cause (PCC). It derives a predictability inequality from exact NMF/PCC and uses its empirical satisfaction to define an effective rank R_c, claiming this estimate is stable against weak noise in contrast to BIC. The paper further claims that NMF at ranks around R_c produces basis images that are stable against noise and against the seed of local optimization, thereby effectively resolving the NMF nonidentifiability problem. Additional contributions include an approximate-PCC interpretation with correlation inequalities and entropy relations, a clustering method based on common-cause weights, and denoising experiments comparing NMF with PCA.","tokens_in":18471,"tokens_out":9016,"duration_ms":96061,"significance":"If the central claims were fully substantiated, the paper would offer a practical rank-selection criterion and a resolution of NMF nonidentifiability, which are longstanding practical problems. The exact derivation of inequality (16) is clean, and the conceptual link between NMF and PCC is interesting and potentially fruitful. The BIC comparison and the denoising experiments, including the random-accuracy baseline, are useful additions. However, the current evidence for the strongest claims is incomplete: the noise-stability experiment is confounded by a data split, no chance baseline is provided for the matched cosine distances, and the paper itself acknowledges missing semantic analysis. The significance is therefore conditional on substantial revision.","major_comments":[{"comment":"The noise-stability experiment is confounded: the clean and noisy NMF runs are performed on two different halves of UTK, so any distance between the matched basis images reflects not only noise but also the difference between two independent subsets of data. This cannot establish 'stability against noise'. In addition, Table V reports maximum matched distances of 0.9175–0.9901 at R=36, and no null model is given for what the distance distribution would be under random local minima or random matching. The paper's own Section IX states that semantic analysis tools are lacking. The abstract's claim that nonidentifiability is 'effectively resolved' is therefore not supported by the presented evidence.","section":"Section VI"},{"comment":"There is an internal inconsistency in the noise-stability claim for the rank estimator. The text states 'Rc(0.05) = 46 > Rc(0) = 30' for UTK, but Table II reports for UTK with ξ=0.05 the values R=22 (τ=10^{-4}) and R=28 (τ=10^{-5}). Since the relation between Rc(ξ) and R(ξ,τ) is not defined, the reader cannot verify whether Rc increases or decreases with noise. The definition of Rc for noisy data needs to be made precise, and the seed variability of Rc should be reported.","section":"Section IV-A2 and Table II"},{"comment":"The inequality S_i ≤ S-hat_i is stated as a result but no proof or empirical support is given. This inequality is not true for arbitrary approximations; for a mixture, conditional entropy can increase or decrease depending on the component distributions and weights. If it is meant to hold for NMF local minima, a proof or a precise statement of the conditions is needed. If it is an empirical observation, the supporting experiments should be shown. As written, it is an unsupported formal claim in the section that motivates the approximate-PCC interpretation.","section":"Section V-B, Eq. (23)"},{"comment":"The paper itself limits the central claim: it states that different local minima 'do not necessarily mean that these different basis images refer to different semantic features' and that semantic analysis tools are still under development. This directly contradicts the abstract and Section VI, which present the nonidentifiability problem as effectively resolved. To support 'effective resolution', the authors need either a semantic evaluation of the outlier basis-image pairs or a quantitative bound on the proportion of unstable basis images. The current mean/median matched distances are not sufficient.","section":"Section IX"}],"minor_comments":[{"comment":"The dataset dimensions are inconsistent with Section III-A: Swimmer is given as (M=256, N=132) although images are 13x13=169 pixels; Olivetti is written as (M=400, M=642) with a duplicated M; UTK is given as (M=1500, N=502) whereas the text says 50x50=2500 pixels. These notation errors should be corrected.","section":"Table II caption"},{"comment":"The caption of Figure 4 reports mean and median matched cosine distances of 0.0236 and 0.0615 for R=36, while Table V for Fig. 5 at R=36 gives mean 0.1056 and median 0.0638. The captions also differ in whether the two halves were trained with the same or different seeds. This discrepancy needs clarification.","section":"Figure 4 caption vs Table V"},{"comment":"The phrase 'the fraction of valid inequalities converges to 1' is vague; the threshold used to define R_c should be stated explicitly. The notation R_c, R(ξ,τ), and R_c(ξ) should be unified across the text and tables.","section":"Section IV-A"},{"comment":"The histograms in Figures 5 and 6 and the summary statistics in Table V would be more informative if the fraction of pairs with distance above a threshold (e.g., 0.5) were reported, together with a null distribution obtained by matching basis images from unrelated NMF runs or from random nonnegative matrices.","section":"Section VI"}],"recommendation":"major_revision","confidential_remarks":"The paper contains a strong conceptual idea and several useful empirical observations, but the central stability claim is currently overclaimed relative to the evidence. The main technical fixes—deconfounding the noise experiment, adding null baselines, clarifying the rank-estimator definitions, and proving or relabeling Eq. (23)—are feasible within the scope of the manuscript. I would not recommend rejection, but the revision needs to be substantial."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the predictability-inequality rank heuristic is a nice idea, and the approximate-PCC observations are genuinely interesting, but the paper's central claim—that NMF around R_c 'effectively resolves' nonidentifiability—goes beyond what the evidence supports. The stability analysis is too thin to carry that weight.\n\nWhat's new: using (16) as a rank-selection criterion is a legitimate extension of a simple result, and the empirical behavior across Swimmer, Olivetti, and UTK is plausible. The anticorrelations in Tables III and IV are real empirical observations, and the denoising comparison is a useful sanity check. I also credit the authors for flagging the outlier-pixel effect and the nonmonotonic seed-stability behavior; they're not hiding warts.\n\nSoft spots, in rough proportion: (1) The stability evidence. One train/test split, one seed pair per condition, no chance baseline. Hungarian matching will always produce some low distances, and the observed long tails—max matched distance 0.99 at R=36 in the seed-only condition—undercut 'effectively resolved.' The authors' own Section IX says they lack semantic analysis to know whether distant matches are truly different, which is exactly the missing experiment. (2) The entropy inequality (23) is stated without proof and, as written, is not generally true for arbitrary approximations; it needs either a proof under the NMF constraints or a counterexample. (3) The BIC comparison is suggestive but doesn't include other modern rank-selection methods, so the comparative claim is thin. (4) R_c itself is defined at local minima; the 10-seed check is okay but doesn't quantify variability of the criterion.\n\nOn balance: the central argument is a heuristic, not a theorem. The paper is worth serious refereeing because the heuristic is new and potentially useful, and the approximate-PCC angle might spark more work. I would bring it to a reading group, but I would not cite the nonidentifiability-resolution claim in its current form. Recommend major revision: add chance baselines, repeated splits, a proof or precise statement for (23), and dial back the abstract.","headline":"A promising rank-selection heuristic for NMF, but the evidence doesn't support the 'resolves nonidentifiability' claim.","tokens_in":18935,"tokens_out":5102,"would_cite":false,"duration_ms":49077,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The principle of the common cause picks the rank where NMF features become stable.","keywords":["nonnegative matrix factorization","principle of the common cause","effective rank selection","basis image stability","nonidentifiability","conditional independence","clustering","image denoising"],"falsifier":"Run the same stability protocol on a synthetic dataset with known ground-truth nonnegative rank and known generating parts: if basis images around Rc fail to recover those parts under the paper's noise model, or if the matched cosine distances are no better than distances obtained by matching random basis-image pairs, the rank/stability claim would be refuted.","tokens_in":1637,"feed_emoji":"🧩","tokens_out":1968,"duration_ms":66010,"temperature":0.7,"pith_summary":"Nonnegative matrix factorization and the principle of the common cause turn out to describe the same operation: explaining a matrix of probabilities through a small set of hidden causes. The paper uses a predictability inequality inherited from the common-cause principle to choose the effective rank of NMF, and shows that basis images learned at that rank are stable under noise and under changes of the random seed—so the usual nonidentifiability of NMF does not block interpretable feature extraction. Conversely, NMF supplies an approximate version of the common-cause principle: larger and positively correlated joint probabilities are explained better by the independent mixture, which yields a clustering rule and a denoising procedure. If the stability claim holds, the practical takeaway is that a data-driven rank choice can make NMF features meaningful despite the factorization's known non-uniqueness.","feed_headline":"Common-cause rule sets NMF rank that survives noise","feed_subtitle":"Rank chosen by common-cause predictability keeps NMF basis images stable under noise and seed changes.","key_machinery":"The engine is the predictability inequality (16), proved from exact conditional independence: for each pair of pixel and image there exist basis images b1 and b2 such that p(π|b1) ≤ p(π|i) ≤ p(π|b2). The smallest rank at which this inequality holds for essentially all pixels defines the effective rank Rc. Stability is then assessed by comparing two sets of basis images with cosine distances and optimal one-to-one matching; the regime around Rc is where matched distances stay small, while for much larger ranks the basis images become unstable.","core_discovery":"The paper's central claim is that a predictive inequality derived from the principle of the common cause—for every pixel and image, some basis image predicts the pixel at least as strongly as the image itself, in both directions—gives a workable way to choose the effective rank of a nonnegative matrix factorization. Around the resulting rank Rc, the basis images found by local optimization are stable under bit-flip noise and under changes of the random seed, so the nonidentifiability of NMF does not prevent reliable feature extraction. The paper also claims that NMF implements an approximate form of the common-cause principle: the higher and more positively correlated joint probabilities are","pith_inferences":["Beyond the paper: if the stability result transfers to other nonnegative data matrices (not just gray-scale images), the same rank-selection rule could serve as a default hyperparameter choice for NMF in applications where basis interpretability matters.","Beyond the paper: the approximate common-cause principle suggests a testable design criterion for NMF objectives—explicitly upweighting large positive joint probabilities should reproduce or improve the observed anticorrelation between approximation error and probability size.","Beyond the paper: the nonmonotonic distance distribution seen when changing seeds at different ranks is an open clue; a systematic study with many seeds and chance-level baselines could show whether the apparent stability around Rc is a true geometric property or a selection effect.","Beyond the paper: because the rank estimate is derived from a conditional-independence condition, it may also apply to other latent-variable models beyond matrix factorization, such as topic models or nonnegative tensor factorizations."],"forward_implications":["Around the common-cause rank Rc, NMF basis images can be treated as interpretable features despite the factorization's non-uniqueness, because they survive noise and seed changes.","The PCC-based rank estimate remains sensible under weak noise in cases where Bayesian Information Criterion based estimates fail to show a minimum.","NMF gives an approximate common-cause principle: larger, positively correlated probabilities are explained better, which can ground clustering by shared causes.","NMF denoises when the rank lies in a middle interval [R1, R2]; below it no reliable features form, and above it the model overfits noise.","Reconstructed images have per-image entropy at least as large as the original conditional distributions, and basis images become sparser than images once the rank exceeds a small threshold."],"supporting_citations":[{"why":"Supplies the principle of the common cause as an exact conditional-independence explanation of dependence.","marker":"[7]"},{"why":"Supplies the probabilistic-causality setup and the proof of the predictability inequality (16) that defines Rc.","marker":"[8]"},{"why":"Defines NMF, reviews its nonidentifiability and feature interpretability, and frames the problem the paper addresses.","marker":"[3]"},{"why":"Provides the BIC rank-selection formulas used as the unstable baseline comparison.","marker":"[5]"},{"why":"Provides the Swimmer dataset and a classic example of nonidentifiability of exact NMF parts.","marker":"[20]"},{"why":"Proposes a generalized maximum-likelihood principle for NMF nonidentifiability, which the paper builds on.","marker":"[21]"},{"why":"Supplies the alternating least-squares NMF algorithm used for all local optimizations.","marker":"[1]"},{"why":"Gives the normalization-preservation property of KL-based NMF used to map images to probability models.","marker":"[17]"},{"why":"Supplies the face dataset with repeated subjects used for clustering and stability experiments.","marker":"[28]"},{"why":"Supplies the large face dataset used for rank estimation, stability, and denoising experiments.","marker":"[29]"}],"fun_headline_variants":["Common cause pinpoints stable NMF rank under noise","Causality principle picks NMF rank that resists noise","PCC gives NMF rank stable against noise and seeds","Common cause rule selects NMF rank stable to noise","How common cause yields stable NMF rank"],"cache_read_input_tokens":20736,"weakest_assumption_plain":"The stability conclusion rests on treating the few computed local minima (one seed pairing or one noise draw, matched optimally) as representative of all NMF solutions; if those minima are unrepresentative, the claim that nonidentifiability is effectively resolved does not follow.","fun_headline_variants_meta":{"raw":{"variants":["Common cause pinpoints stable NMF rank under noise","Causality principle picks NMF rank that resists noise","PCC gives NMF rank stable against noise and seeds","Common cause rule selects NMF rank stable to noise","How common cause yields stable NMF rank"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000467,"raw_usage":{"total_tokens":2159,"prompt_tokens":731,"completion_tokens":1428,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":475,"completion_tokens_details":{"reasoning_tokens":1365}},"tokens_in":475,"tokens_out":1428,"duration_ms":10336,"temperature":1.0,"reasoning_tokens":1365,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:46:39.780480+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same stability protocol on a synthetic dataset with known ground-truth nonnegative rank and known generating parts: if basis images around Rc fail to recover those parts under the paper's noise model, or if the matched cosine distances are no better than distances obtained by matching random basis-image pairs, the rank/stability claim would be refuted.","supporting_citations":[{"cited_title":"Reichenbach, The direction of time","cited_arxiv_id":null,"evidence_quote":"Supplies the principle of the common cause as an exact conditional-independence explanation of dependence."},{"cited_title":"Suppes, A probabilistic theory of causality","cited_arxiv_id":null,"evidence_quote":"Supplies the probabilistic-causality setup and the proof of the predictability inequality (16) that defines Rc."},{"cited_title":"Gillis, Nonnegative Matrix Factorization","cited_arxiv_id":null,"evidence_quote":"Defines NMF, reviews its nonidentifiability and feature interpretability, and frames the problem the paper addresses."},{"cited_title":"Assessing methods for evaluating the number of components in non-negative matrix factorization,","cited_arxiv_id":null,"evidence_quote":"Provides the BIC rank-selection formulas used as the unstable baseline comparison."},{"cited_title":"When does non-negative matrix factor- ization give a correct decomposition into parts?","cited_arxiv_id":null,"evidence_quote":"Provides the Swimmer dataset and a classic example of nonidentifiability of exact NMF parts."},{"cited_title":"The most likely common cause,","cited_arxiv_id":null,"evidence_quote":"Proposes a generalized maximum-likelihood principle for NMF nonidentifiability, which the paper builds on."},{"cited_title":"Positive matrix factorization: A non-negative factor model with optimal utilization of error estimates of data values,","cited_arxiv_id":null,"evidence_quote":"Supplies the alternating least-squares NMF algorithm used for all local optimizations."},{"cited_title":"Non-negative matrix factorization with fixed row and column sums,","cited_arxiv_id":null,"evidence_quote":"Gives the normalization-preservation property of KL-based NMF used to map images to probability models."},{"cited_title":"The database of faces,","cited_arxiv_id":null,"evidence_quote":"Supplies the face dataset with repeated subjects used for clustering and stability experiments."},{"cited_title":"Age progression/regression by condi- tional adversarial autoencoder,","cited_arxiv_id":null,"evidence_quote":"Supplies the large face dataset used for rank estimation, stability, and denoising experiments."}],"review_version":1}