{"id":"433d275e-1e1d-4a73-9675-e6d2e0b00290","arxiv_id":"2501.18028","paper_version":3,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A rank-conscious Gini prametric yields competitive KNN and K-means benchmarks on 16 UCI datasets, but the central convergence proof and the fair-evaluation protocol do not hold up.","lead":"The paper proposes a distance that mixes value differences with rank differences, called a Gini prametric, and uses it in K-means and k-nearest neighbor classifiers. The reported results show it competitive with the Hassanat distance on 16 UCI datasets, but the proof that Gini K-means converges is mathematically unsound.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Convergence proof of Gini K-means rests on a false centroid-minimizer lemma; the arithmetic mean is not the minimizer of the within-cluster Gini objective.","rationale":"The reader's weakest assumption identifies exactly the load-bearing flaw: the convergence proof of Proposition 4 depends on the false claim that the arithmetic mean minimizes the within-cluster sum of squared Gini distances when ranks are fixed. I re-examined the supplementary proof and found the algebraic step invalid: the derivative equation is a single scalar condition that cannot determine a d-dimensional centroid, and even in one dimension it resolves to a weighted mean whose weights depend on rank differences, not to the simple average. A concrete scalar counterexample with points 0, 1, 4 and nu = 2 shows the minimizer is 5/11, not 5/3. Because every iteration of Lloyd-style K-means must reduce the objective for convergence, and the proof's inequality (1) relies on this centroid-minimizer lemma, the convergence theorem is unsupported. The empirical concerns about oracle selection of nu in the K-means experiments are secondary but reinforce the rejection; the KNN results are more credible but do not rescue the paper's central theoretical claim. No ad hominem is intended; the issue is a concrete mathematical error. The verdict should remain REJECT, so no change to the reader's verdict is needed.","tokens_in":43628,"tokens_out":4366,"duration_ms":54535,"concrete_test":"Re-derive the centroid update for a one-dimensional cluster under the same frozen-rank assumption used in the proof. Take C = {0, 1, 4}, nu = 2, and treat R(z) = 2.5 as constant on the interval (1, 4). Compute f(z) = sum_i (x_i - z)^2 (R(x_i) - R(z))^2 = 2.25 z^2 + 0.25(1 - z)^2 + 0.25(4 - z)^2. Minimizing gives z = 5/11, not the arithmetic mean 5/3. This directly falsifies the proof's step 'This implies z = 1/n sum_i z_i'. A broader check would verify whether Lloyd-style updates with arithmetic-mean centroids actually decrease the Gini objective on this example; they do not, since the mean lies outside the frozen-rank interval where the update is derived.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The convergence proof of Proposition 4 (Supplementary Materials, Proof of Proposition 4) asserts that the arithmetic mean is the unique minimizer of the within-cluster sum of squared Gini distances when ranks are held fixed. This is false. From the derivative condition, the proof derives a single scalar equation, sum_i (z_i - z) dot (R_Z(z_i)^{nu-1} - R_Z(z)^{nu-1}) = 0, and then concludes z = (1/n) sum_i z_i. Even granting that equation, it cannot determine a vector z in R^d; in the scalar case it yields a weighted mean with weights w_i = (R(z_i)^{nu-1} - R(z)^{nu-1})^2, not the unweighted arithmetic mean. For example, with cluster {0, 1, 4}, nu = 2, and the rank of z held fixed at 2.5, the objective is minimized at z = 5/11 approximately 0.4545, not at the arithmetic mean 5/3. Since inequality (1) in the proof relies entirely on this centroid-minimizer lemma, the claimed convergence of Gini K-means is not established. This is the load-bearing theoretical claim of the paper; without it, the central novelty of a provably convergent rank-based K-means collapses.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper defines a generalized Gini prametric dG,ν that combines value differences with rank differences between points, and uses it to build variants of K-nearest neighbors and K-means. The main claims are: (i) the Gini prametric is robust to noise and outliers; (ii) Gini K-means is provably convergent when rank vectors remain constant; and (iii) Gini KNN is competitive with Hassanat distance. The theoretical convergence claim is delegated to the Supplementary Materials, and the empirical evaluation is carried out on 16 UCI datasets with and without added noise, reporting precision and recall rankings for KNN, K-means, and agglomerative clustering.","tokens_in":43842,"tokens_out":8296,"duration_ms":323620,"significance":"If the convergence theorem were correct, the paper would offer a practical, provably convergent rank-based alternative for robust K-means and KNN, and the extensive UCI experiments would be a useful benchmark. The paper also ships supplementary tables and describes reproducible experimental machinery. However, the central convergence proof rests on an incorrect centroid-update lemma, and the headline K-means experiments select the hyperparameter ν using labeled test data. These two issues undermine the paper's main theoretical and empirical contributions as they currently stand.","major_comments":[{"comment":"The proof's key lemma—that the arithmetic mean uniquely minimizes the within-cluster sum of squared Gini distances under fixed ranks—is false. The derivative equation shown in the proof, even if accepted, gives a weighted mean, not the unweighted arithmetic mean. Concretely, for the scalar cluster {0,1,4} with ν=2 and descending ranks (3,2,1), for any centroid z between 1 and 4 the objective is f(z)=z^2+(4-z)^2, whose minimizer is z=2, not the arithmetic mean 5/3. In addition, the differentiation step itself is algebraically incorrect: the squared distance is the square of a sum over j, so differentiating Σ_i [Σ_j (z_ij−z_j)(R_ij^{ν−1}−R(z_j)^{ν−1})]^2 does not produce the simple equation written in the proof. Since inequality (1) depends entirely on this centroid-minimizer lemma, Proposition 4 is not established.","section":"Supplementary Materials, Proof of Proposition 4"},{"comment":"The convergence statement is conditional on 'rank vectors stay constant', but the algorithm does not enforce or verify this condition. The centroid is updated to the arithmetic mean of the cluster, and there is no reason that the ranks of the new centroid with respect to the cluster members remain unchanged. Thus, even if the centroid-minimizer lemma were correct, the proof would cover only a hypothetical fixed-rank iteration, not the algorithm actually implemented and evaluated in the experiments.","section":"Section 4.2, Proposition 4 and Algorithm 1"},{"comment":"The reported K-means advantage is obtained by selecting ν to maximize precision on the supervised UCI datasets, i.e., using the test labels. The text states: 'a grid search is performed over the hyper-parameter ν, selecting the value that maximizes the precision score.' This is an oracle selection that inflates the mean ranks in Tables 6 and 7. The paper's own later silhouette-based selection (Tables 8 and 9) drops Gini K-means to rank 4 in precision and rank 3 in recall, far below the headline ranks of 3.50 and 4.19. The empirical claim of superiority therefore depends on information that would not be available in a real unsupervised application.","section":"Section 5.2, K-means Experiments"},{"comment":"The statement that the Cover-Hart asymptotic error bound applies unchanged to the generalized Gini prametric is not justified. Cover and Hart's result relies on the metric structure of the space, in particular the property that the nearest neighbor converges to the query point as the sample size grows. The Gini prametric does not satisfy the triangle inequality and can assign zero distance to distinct points with identical rank vectors, so the standard proof does not carry over. At minimum, a proof is needed; the sentence 'The proof is exactly the same as the one of Cover [9]' does not address these differences.","section":"Section 4.1, Proposition 3"},{"comment":"There is an internal inconsistency about the domain of ν. Definition 2 and the surrounding text require ν>1, and this condition is what guarantees non-negativity of dG,ν. However, Algorithm 1 searches over ν in [0.1;6] excluding 1, and Proposition 4 claims convergence for all ν≠1. For 0<ν<1, the exponent ν−1 is negative, so the generalized Gini prametric can be negative and is no longer a prametric in the sense used elsewhere in the paper. The grid search and the convergence statement should be restricted to the range where the distance is well defined.","section":"Definitions 2 and Algorithm 1"}],"minor_comments":[{"comment":"The text says 'it can be referred to as a Gini pramatric space'; the intended term is 'prametric space'.","section":"Section 3.3"},{"comment":"Reference [9] appears to be a duplicate of [10] and lists the author as 'H. Cover'; the correct citation is T. M. Cover and P. E. Hart, 'Nearest neighbor pattern classification', IEEE Transactions on Information Theory, 1967.","section":"References [9] and [10]"},{"comment":"The line 'RX = conditional ranks of X_tr[-fold]' is unclear: the notation X_tr[-fold] is not defined, and the procedure for computing conditional ranks when a fold is held out should be specified precisely.","section":"Algorithm 1"},{"comment":"The phrase 'The KNN algorithm converges' is not defined; KNN is a classification rule without an iterative update, so the statement should be replaced with a precise consistency or asymptotic-error claim.","section":"Section 4.1, Proposition 3"}],"recommendation":"reject","confidential_remarks":"The paper cites its own arXiv preprint [27] as the source of the proof of Proposition 4; this is a self-citation and should be flagged internally. The experimental protocol in Section 5.2 selects ν using test labels, which is a serious methodological issue that the authors may be able to fix in a revision, but the false centroid-minimizer lemma in the Supplementary Materials is a load-bearing mathematical error that undermines the main theoretical claim. Combined with the unjustified use of the Cover-Hart bound in Proposition 3, I do not see how the current manuscript can be accepted without a fundamentally different convergence argument."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. The Gini prametric is a genuinely new construction: it multiplies value differences by rank differences and throws in a tunable exponent ν, and that is a coherent way to build a robust, non-metric similarity for KNN and K-means. The KNN experiments are the solid half of the paper. The other half—the claimed convergence proof for Gini K-means and the headline clustering gains—does not hold up.\n\nWhat the paper does well: the distance definition is natural, the invariance properties are stated cleanly, and the experimental comparison is broad (16 UCI data sets, 12 distances, noise at 5% and 10%). The authors also report a Wilcoxon test. Those p-values show no significant advantage over Hassanat's distance, which is honest evidence even though it directly contradicts the abstract's 'superior performance' claim. They also disclose that the K-means ν selection in the main tables uses the true labels, and they include a silhouette-based selection experiment that drops the method to roughly fourth place. That transparency is to their credit.\n\nThe soft spots are load-bearing. The supplementary proof of Proposition 4 is algebraically wrong. From the derivative condition, the paper concludes that the centroid must be the arithmetic mean. In the scalar case the stationarity equation gives a weighted mean with weights (R(z_i)^{ν−1} − R(z)^{ν−1})², not the unweighted average. For the cluster {0, 1, 4} with ν = 2, the minimizer is about 0.45, not 5/3. And the per-point inequality at the start of the proof is not a consequence of the assignment step: a point can be farther from the updated centroid than from the previous one. So the central convergence theorem is not established. On the empirical side, selecting ν by maximizing precision on the test labels makes Tables 6–7 oracle results; the silhouette-based Tables 8–9 are the honest comparison and they put Gini mid-pack. The KNN results are more credible, but even there the hyperparameters are tuned on the metric being reported, and the Wilcoxon test undercuts the 'superior' language.\n\nWho is this for? Someone working on rank-based similarities or robust KNN might find the distance worth trying, and the paper is a reasonable pointer to that literature. But the math needs to be corrected or removed, and the clustering evaluation needs to be rerun without label peeking. I would not desk-reject it; an expert referee could check the proof and force a major revision. As submitted, the central claims are not supported.","headline":"Gini distance is worth a look, but the convergence proof is wrong and the K-means win is an oracle result.","tokens_in":44417,"tokens_out":5971,"would_cite":false,"duration_ms":66372,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30","68T10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Replacing metric distances with a Gini prametric—a value-rank hybrid—is claimed to make K-means and KNN robust to noise, with Gini K-means provably convergent when rank vectors stay constant.","keywords":["Gini prametric","K-means","K-nearest neighbors","rank-based distance","robust clustering","prametric space","noise robustness","generalized Gini index"],"falsifier":"Evaluate the within-cluster objective $f(z)=\\sum_i d_{G,\\nu}(x_i,z)^2$ for a small one-dimensional cluster such as $\\{0,2,10\\}$ with fixed ranks $(1,2,3)$, and compare $z=4$ (the arithmetic mean) with a weighted mean derived from the first-order condition. If the weighted mean gives a strictly smaller value of $f$, the centroid-update lemma behind Proposition 4 is false; conversely, if the arithmetic mean always wins, the convergence proof is on solid ground.","tokens_in":43351,"feed_emoji":"📊","tokens_out":8787,"duration_ms":77553,"temperature":0.7,"pith_summary":"The paper proposes replacing standard metric distances in K-means and K-nearest neighbors with a Gini prametric, a symmetric non-distance that multiplies coordinate differences by differences of powered ranks. It claims this rank-plus-value measure is robust to noise and outliers, and that the resulting Gini K-means converges whenever rank vectors stay constant for any $\\nu \\neq 1$. In experiments on sixteen public benchmark datasets, the generalized Gini KNN ranks at or near the top in precision and recall, matching or exceeding the Hassanat distance, which the paper treats as the strongest noise-robust baseline. If these claims hold, rank-based prametrics would be a practical drop-in replacement for metric distances in clustering and classification, giving provable convergence and noise resistance without the triangle inequality.","feed_headline":"Rank-aware Gini distance beats noise in K-means and KNN","feed_subtitle":"Non-metric distance combining value and rank gaps matches the top robust baseline on 16 datasets.","key_machinery":"The load-bearing object is the generalized Gini prametric $d_{G,\\nu}$, a value-rank hybrid in which each coordinate gap $(x_{i,j}-x_{k,j})$ is weighted by the gap between powered descending ranks $R_{X_j}(\\cdot)^{\\nu-1}$. The hyperparameter $\\nu$ tunes sensitivity to the tail of the feature distribution: $\\nu=2$ treats upper and lower ranks symmetrically, while larger $\\nu$ emphasizes higher ranks and smaller $\\nu$ emphasizes lower ranks. This object carries the argument because it injects robustness through ranks while still reacting to actual value differences, and its invariance properties (linear and rank invariance) are what the convergence and robustness arguments rely on.","core_discovery":"The central claim is that the generalized Gini prametric $d_{G,\\nu}(x_i,x_k) = -\\sum_{j=1}^d (x_{i,j}-x_{k,j})(R_{X_j}(x_{i,j})^{\\nu-1} - R_{X_j}(x_{k,j})^{\\nu-1})$, where $R_{X_j}$ is the descending rank of feature $j$, is a valid dissimilarity for KNN and K-means even though it is not a metric. The paper proves in Proposition 4 that Gini K-means converges for all $\\nu \\neq 1$ when rank vectors stay constant, and reports that Gini KNN achieves the best average precision rank among twelve distances on clean data, with competitive performance under added Gaussian noise. The same construction is extended to agglomerative hierarchical clustering, where the generalized Gini prametric ranks first in precision on noisy data.","pith_inferences":["A natural repair if the centroid lemma fails is to replace the arithmetic mean by the weighted mean that the derivative condition actually selects, and to prove a Lloyd-style potential decrease for that update; this would salvage convergence for a modified Gini K-means.","The conditional-rank construction used for train-test splits suggests that streaming or online use of the Gini prametric would need an explicit rule for updating ranks as new points arrive, since distances between old points change under re-ranking.","One could test whether $\\nu$ should be adapted to the noise level itself: the reported advantage over the Hassanat distance is clearest at 5% noise and shrinks at 10%, so a noise-dependent schedule for $\\nu$ may extend the method's robustness.","The same value-rank weighting idea could be applied to other algorithms that only need pairwise dissimilarities, such as spectral clustering or distance-based outlier detection, though the paper only demonstrates K-means, KNN, and agglomerative clustering."],"forward_implications":["If Proposition 4 is correct, Gini K-means is guaranteed to stop for any $\\nu\\neq 1$ whenever the rank vectors of the data do not change during iteration, making rank-based robust clustering a provably convergent procedure.","Gini KNN inherits the standard nearest-neighbor asymptotic error bound (Proposition 3), so the rank-based distance adds robustness without sacrificing the classical consistency guarantee.","The hyperparameter $\\nu$ gives a tunable robustness lever: practitioners can shift weight between lower and upper parts of each feature's distribution, with $\\nu=2$ as the symmetric baseline.","Because the prametric is linear-invariant and rank-invariant, monotone transformations of the features leave the distances essentially unchanged, which the paper argues protects against outliers and measurement errors."],"supporting_citations":[{"why":"Supplies the 44-distance benchmark and identifies the Hassanat distance as the strongest noise-robust baseline the paper compares against.","marker":"[1]"},{"why":"Defines prametric (premetric) spaces, the topological setting that permits dropping symmetry and triangle inequality.","marker":"[3]"},{"why":"Introduces the Gini mean difference, the value-plus-rank dispersion measure the prametric generalizes.","marker":"[16]"},{"why":"Defines the Gini covariance operator used to construct the Gini prametric from value-rank products.","marker":"[32]"},{"why":"Introduces the extended Gini index with parameter ν, the basis for the generalized prametric.","marker":"[33]"},{"why":"Provides the Gini methodology properties (invariance, covariance identities) invoked in Propositions 1 and 2 and in the convergence proof.","marker":"[45]"},{"why":"Gives the nearest-neighbor convergence and error-bound proof that Proposition 3 says applies unchanged to Gini KNN.","marker":"[9]"},{"why":"Contains the proof of Proposition 4, the convergence theorem for Gini K-means.","marker":"[27]"},{"why":"Defines the silhouette score used in Algorithm 1 to select the hyperparameter ν in unsupervised settings.","marker":"[31]"}],"fun_headline_variants":["Gini prametric distance strengthens K-means and KNN against noise","Rank-aware Gini distance proves superior for K-means and KNN","Combining rank and value gaps: Gini distance for robust ML","Non-metric Gini prametric rivals top distance on noisy benchmarks","Gini K-means and KNN: rank-based metrics beat noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The convergence theorem depends on the lemma that, with ranks held fixed, the arithmetic mean of a cluster is the unique minimizer of the sum of squared Gini distances to the cluster points; if the true minimizer is instead a weighted mean, the centroid update in the proof no longer matches the average and the convergence argument collapses.","fun_headline_variants_meta":{"raw":{"variants":["Gini prametric distance strengthens K-means and KNN against noise","Rank-aware Gini distance proves superior for K-means and KNN","Combining rank and value gaps: Gini distance for robust ML","Non-metric Gini prametric rivals top distance on noisy benchmarks","Gini K-means and KNN: rank-based metrics beat noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1420,"prompt_tokens":886,"completion_tokens":534,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":438}},"tokens_in":502,"tokens_out":534,"duration_ms":7393,"temperature":1.0,"reasoning_tokens":438,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T00:58:37.274404+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the within-cluster objective $f(z)=\\sum_i d_{G,\\nu}(x_i,z)^2$ for a small one-dimensional cluster such as $\\{0,2,10\\}$ with fixed ranks $(1,2,3)$, and compare $z=4$ (the arithmetic mean) with a weighted mean derived from the first-order condition. If the weighted mean gives a strictly smaller value of $f$, the centroid-update lemma behind Proposition 4 is false; conversely, if the arithmetic mean always wins, the convergence proof is on solid ground.","supporting_citations":[{"cited_title":"Arkhangelskii and L","cited_arxiv_id":null,"evidence_quote":"Defines prametric (premetric) spaces, the topological setting that permits dropping symmetry and triangle inequality."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Gini mean difference, the value-plus-rank dispersion measure the prametric generalizes."},{"cited_title":"Schechtman and S","cited_arxiv_id":null,"evidence_quote":"Defines the Gini covariance operator used to construct the Gini prametric from value-rank products."},{"cited_title":"Schechtman and S","cited_arxiv_id":null,"evidence_quote":"Introduces the extended Gini index with parameter ν, the basis for the generalized prametric."},{"cited_title":"Yitzhaki and E","cited_arxiv_id":null,"evidence_quote":"Provides the Gini methodology properties (invariance, covariance identities) invoked in Propositions 1 and 2 and in the convergence proof."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the nearest-neighbor convergence and error-bound proof that Proposition 3 says applies unchanged to Gini KNN."}],"review_version":1}