{"id":"a8781953-1d16-4a48-ad2c-f556c5e244ea","arxiv_id":"2411.10418","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A new time series distance, Multiscale Dubuc Distance, compares envelope overlap at multiple scales and is claimed to match tuned DTW in 1-NN classification.","lead":"This paper introduces a new way to measure how similar two time series are, by comparing the up-and-down envelopes around each series at multiple zoom levels. A smart generalist might read it because similarity measures like this are the engine behind time series classification in finance, health care, and space weather, and a faster metric could matter.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 5 is unnormalized: for identical series with E={1,2,4,...,64}, MDS is the trapezoid area 63, so MDD(x,x)=1-63=-62, violating reflexivity and non-negativity.","rationale":"The reader's weakest assumption identifies exactly the load-bearing flaw: Eq. 5 is unnormalized, so the claimed metric properties do not follow and are in fact false under the paper's own experimental setting. This is a direct counterexample to reflexivity, not a subtle or regime-dependent issue. The triangle-inequality inheritance is also invalid because the total weight of the per-scale distances exceeds 1. I find no other concern that is more load-bearing than this, since the entire theoretical section rests on the metric claim. The empirical comparison could potentially be salvaged by a normalization fix, but the paper as written does not provide one, so the rejection verdict remains appropriate.","tokens_in":10005,"tokens_out":2180,"duration_ms":22380,"concrete_test":"Run the authors' provided bitbucket code on any UCR dataset with E={1,2,4,...,64} and compute MDD(x,x) for a fixed training series. If the output is -62 or any negative value, reflexivity fails. Analytically, Eq. 5 with all r=1 gives MDS = sum_{i=2}^{|E|} Delta_epsilon_i = 63 for that E. Additionally, inspect whether the code divides the trapezoidal sum by (epsilon_{|E|}-epsilon_1); if it does, the paper must add that denominator to Eq. 5 and revise the proof accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that MDD is a metric (Sec. IV-A). This fails as written because Eq. 5 computes an area under the intersection-ratio curve, not a weighted average. For any identical pair x, r(x,x,epsilon)=1 for every epsilon, so MDS(x,x,E) equals the trapezoidal integral of the constant function 1 over the interval [epsilon_1, epsilon_{|E|}]. Under the experimental choice E={1,2,4,...,64} (Sec. V-A), the increments Delta_epsilon_i sum to 63, giving MDS(x,x)=63 and MDD(x,x)=1-63=-62. This directly contradicts reflexivity (MDD(x,x)=0) and positivity (MDD>=0), both required for a metric. The triangle-inequality argument inherits from the Jaccard distance only if MDD is a convex combination of per-scale Jaccard distances, i.e., if the weights Delta_epsilon_i are nonnegative and sum to 1. They do not; with sum 63, MDD is not such a combination, so the cited Jaccard-distance result [14,17] does not apply. This is an internal inconsistency in the proof, not merely a mismatch with external consensus. The empirical section does not mention negative distances; if the implementation implicitly normalizes the area (e.g., divides by epsilon_{|E|}-epsilon_1), then Eq. 5 and the metric proof still need that normalization stated explicitly. As submitted, the central theoretical contribution is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Multiscale Dubuc Similarity (MDS), defined in Eq. (5) as the trapezoidal area under the curve of per-scale intersection ratios r(x,y,ε) from Eq. (4), and defines MDD = 1 − MDS. The authors claim that MDD is a metric (or pseudometric), that it runs in linear time, and that 1-NN classification with MDD using a generic scale set E is competitive with DTW with per-dataset optimized windows on 95 UCR datasets. They also present a case study on the Worms dataset where customizing E improves accuracy from 45% to 63%.","tokens_in":10330,"tokens_out":2677,"duration_ms":30833,"significance":"If the metric claim and the empirical comparisons held, MDD would be an interesting addition to the time-series distance toolbox: it is self-contained (no learned constants enter the per-scale ratios), it has a clear multiscale interpretation via Dubuc envelopes, and the paper reports a broad benchmark against EuD, DTW, LCSS, and TS-MIoU. The authors also provide source code and follow a standard 1-NN evaluation framework. However, the central theoretical contribution—that MDD is a metric—is not supported as written, and the reported Worms gain is obtained by selecting E after seeing test-set results, so the paper's headline claims need substantive repair.","major_comments":[{"comment":"The reflexivity and non-negativity claims are contradicted by the unnormalized definition of MDS. For x = y, r(x,x,ε)=1 for every ε, so MDS is the trapezoidal integral of the constant function 1 over [ε_1, ε_m]. With the experimental choice E={1,2,4,...,64} (Sec. V-A), the increments Δε_i sum to 63, giving MDS(x,x)=63 and MDD(x,x)=1−63=−62. This violates the required MDD(x,x)=0 and MDD≥0. The proof that positivity follows from non-negativity of ∩ε and ∪ε is therefore invalid: those non-negativities say nothing about whether MDS exceeds 1. The definition in Eq. (5) needs an explicit normalization (e.g., division by ε_m−ε_1), and all metric proofs and experiments must be re-examined under that normalization.","section":"Sec. IV-A, Eq. (5)"},{"comment":"The triangle-inequality argument is not established. The paper states that MDD inherits the triangle inequality from r because the Jaccard distance is a metric [14,17], but r in Eq. (4) is the ratio of two sums over time indices, Σ ∩ε / Σ ∪ε, not the Jaccard similarity of the two envelope regions as sets. The cited results on Jaccard distance do not automatically apply to this ratio-of-sums quantity. Moreover, even if each per-scale r were a Jaccard similarity, MDD = 1 − ∫ r dε is a metric only if the integral is a convex combination of per-scale Jaccard distances, which requires the weights Δε_i to be nonnegative and sum to 1. They do not in the current formulation, so the stated inheritance from [14,17] is unsupported.","section":"Sec. IV-A, Eq. (4)"},{"comment":"The Worms customization result is circular and does not support the claim that MDD's single parameter is a useful noise-sensitivity gauge. The authors report that with the generic E the accuracy was 45%, and that after 'taking a closer look' they redefined E to {16,32,...,256}, boosting accuracy to 63%. This selection is made after observing test-set accuracy on the same dataset, so the 40% improvement over generic E and the 18% improvement over DTW are not out-of-sample findings. A proper evaluation would preselect E on the training set (or via nested cross-validation) and then report test accuracy.","section":"Sec. V-B"},{"comment":"The linear-time claim is overstated. The pseudocode and Eq. (3) give Θ(|E|·d) per distance computation, and the paper argues that |E| is a user-fixed constant that does not scale with problem size. However, in the experimental setup E is chosen as powers of two up to α·d, so |E| = O(log d). Under that protocol the complexity is Θ(d log d), not Θ(d). If the authors intend |E| to be a true constant independent of d, that should be stated and the experimental protocol should not define E as a function of series length.","section":"Sec. IV-B"}],"minor_comments":[{"comment":"The caption says MDS returns a value between 0 and 1, but Eq. (5) as written is an unnormalized area and can exceed 1; the caption should be updated to match the corrected definition.","section":"Fig. 1 caption"},{"comment":"There are several typographical errors, including 'Multisclae' in the Fig. 1 caption, 'Mulstiscale' in Sec. IV-A, and the source-code URL 'multiscale dubuc' containing a space.","section":"Throughout"},{"comment":"The notation in the MDD definition is inconsistent: MDD(x,y,ε) is written with a scalar ε, but the definition in Eq. (5) depends on the whole set E; the authors should write MDD(x,y,E).","section":"Sec. IV-A"},{"comment":"The text says LCSS points were removed from the Texas Sharpshooter plot for visibility, but the corresponding accuracy and gain values are not reported; a supplementary table or histogram would make the comparison reproducible.","section":"Sec. V-B"}],"recommendation":"reject","confidential_remarks":"The central theorem is false as stated because the unnormalized area in Eq. (5) makes MDD negative for identical series under the paper's own experimental E. The triangle-inequality proof is also not a straightforward application of Jaccard-distance results. The empirical benchmark could potentially be salvaged with a normalized definition and a non-circular Worms experiment, but those changes go beyond local revision. The paper is likely better resubmitted as a corrected manuscript rather than revised in place."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new idea here is the combination of Dubuc's envelope variation with pointwise interval IoU, aggregated over scales. That is genuinely fresh, and the 95-dataset UCR benchmark against DTW, EuD, LCSS, and TS-MIoU is a real piece of work. The algorithm is simple, linear-time, and the paper is clearly written for the most part. If the distance function were properly normalized, this could be a useful contribution to the time-series similarity toolbox.\n\nThe soft spot is load-bearing. Eq. 5 defines MDS as the trapezoidal area under the intersection-ratio curve. For any identical pair, r(x,x,eps)=1 for every eps, so MDS(x,x,E) equals the length of the interval spanned by E. With the experimental choice E={1,2,4,...,64}, that length is 63, so MDD(x,x)=1-63=-62. That violates reflexivity and non-negativity, both of which are required for a metric. The proof in Sec. IV-A inherits the Jaccard triangle inequality only if the weights Delta_epsilon_i sum to 1; they do not, so the cited results do not apply. The authors seem to have intended a weighted average but wrote an area. This is an internal inconsistency in the central claim, not a matter of interpretation.\n\nThere is a second, smaller issue: the Worms demonstration selects E after seeing test-set accuracy. That makes the reported 40% improvement partly circular. No significance testing is reported on the aggregate numbers, so the claim of being 'comparable to optimized DTW' is suggestive but not conclusive.\n\nWhat is solid: the per-scale IoU definition is clean, the time complexity argument is correct, and the empirical comparison is broad. The paper is honest about its proof-of-concept status and lists datasets where MDD underperforms.\n\nThe metric proof can likely be repaired by normalizing Eq. 5, e.g., dividing by the sum of the Delta_epsilon_i or restricting E to consecutive integers. But as submitted, the main theoretical contribution is unsupported.\n\nMy recommendation: send this to peer review, not desk reject. The kernel is novel and the empirical sweep is useful, but the authors must fix the normalization and clarify the parameter selection. A serious referee can guide that revision. I would not cite the current version in my own work because the metric property is central and currently wrong.","headline":"The construction and benchmark are worth a look, but the claim that MDD is a metric fails as written because Eq. 5 computes an unnormalized area, not a convex combination.","tokens_in":10875,"tokens_out":2155,"would_cite":false,"duration_ms":20827,"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":"This paper claims that Multiscale Dubuc Distance, a multiscale envelope-overlap measure, is a metric that matches tuned DTW's 1-NN accuracy on 95 UCR datasets while running in linear time.","keywords":["time series","similarity measure","Dubuc variation","Jaccard index","intersection over union","metric","dynamic time warping","1-NN classification"],"falsifier":"Compute Eq. 5 for a series paired with itself under the paper's generic scale set; with $E=\\{1,2,4,\\ldots,64\\}$ the area under the curve is 63, so $\\mathrm{MDD}(x,x)=1-63=-62$, contradicting the reflexivity axiom that the metric proof claims.","tokens_in":1848,"feed_emoji":"📈","tokens_out":2151,"duration_ms":76381,"temperature":0.7,"pith_summary":"This paper introduces Multiscale Dubuc Distance (MDD), a similarity measure for equal-length time series that combines Dubuc's variation envelopes from fractal analysis with a Jaccard-style intersection-over-union ratio, integrated across multiple scales. The authors claim MDD is a metric satisfying the triangle inequality, and that a 1-NN classifier using MDD with a fixed generic set of scales achieves accuracy comparable to DTW with per-dataset optimized windows on 95 UCR datasets, while running in linear time. If the metric claim holds, MDD would be a fast, tuning-free alternative to DTW for large-scale time-series classification and retrieval. The accuracy and complexity claims are supported by the experiments; the metric proof rests on a scaling assumption that the paper's own default scale set violates.","feed_headline":"Time-series distance matches tuned DTW in linear time","feed_subtitle":"Envelope-overlap metric MDD rivals tuned DTW on 95 UCR datasets in linear time.","key_machinery":"The core object is the Dubuc variation envelope pair at scale $\\varepsilon$: for each time series, define the upper envelope $u_{x,\\varepsilon}(t)$ as the supremum of the series' values within distance $\\varepsilon$ of $t$, and the lower envelope $l_{x,\\varepsilon}(t)$ as the infimum. The intersection $\\cap_\\varepsilon(x,y)$ and union $\\cup_\\varepsilon(x,y)$ are computed pointwise from the two envelopes, and their ratio $r(x,y,\\varepsilon) = \\cap_\\varepsilon / \\cup_\\varepsilon$ is a Jaccard index at that scale. The similarity MDS is the area under the curve obtained by trapezoidal integration of $r$ over the scales in $E$, and $\\mathrm{MDD} = 1 - \\mathrm{MDS}$. The Jaccard ratio is what carries the claimed metric properties, while the area under the curve is what fuses the scales into a single number; the linear-time complexity comes from computing bounds, intersection, and union each in one pass.","core_discovery":"The paper's central claim is that Multiscale Dubuc Distance (MDD) is a metric—more cautiously, a pseudometric—for time series of equal length. The construction borrows Dubuc's variation envelopes: for each scale $\\varepsilon$, each series is replaced by the upper and lower bounds of its values in a sliding neighborhood of radius $\\varepsilon$; the overlap of the two resulting strips is normalized by their union, producing a Jaccard-style intersection ratio $r(x,y,\\varepsilon)$. The similarity MDS is the trapezoidal area under the $r$-versus-$\\varepsilon$ curve, and $\\mathrm{MDD} = 1 - \\mathrm{MDS}$. The metric property, including the triangle inequality, is argued by inheritance from the Jaccard distance at a single scale. Experimentally, on 95 UCR datasets, 1-NN with MDD using a fixed generic set of scales reaches average accuracy statistically comparable to DTW with a per-dataset tuned window, and the algorithm runs in time linear in the series length. The paper also shows that shrinking the smallest scale in $E$ makes MDD more sensitive to noise, and enlarging it suppresses noise, which on datasets like Worms yields a large accuracy gain.","pith_inferences":["The reflexivity failure is fixable: normalizing the area under the curve by the total measure of $E$ (e.g., dividing by $\\sum \\Delta\\varepsilon_i$) restores $\\mathrm{MDD}(x,x)=0$, and the resulting quantity is a convex combination of single-scale Jaccard distances, which would likely inherit the triangle inequality if all scales share the same normalization.","The metric property and classification performance are logically independent; even if the metric proof collapses, MDD may remain a strong 1-NN similarity, so the two claims should be evaluated separately.","A natural testable extension is an automatic or learned choice of the minimal scale, which could make MDD's already competitive accuracy exceed tuned DTW on noise-heavy datasets without manual inspection.","Because the envelopes are defined without an aspect-ratio parameter, the same construction could be applied to other one-dimensional signals or, with a suitable neighborhood, to images and graphs, suggesting MDD-style multiscale overlap is a general dissimilarity template."],"forward_implications":["On the 95 UCR datasets tested, 1-NN with MDD using one generic scale set beats or ties DTW with a learned window on 31% of datasets and matches DTW's average accuracy, so a tuning-free MDD could serve as a drop-in DTW replacement in many pipelines.","Because MDD runs in $\\Theta(d)$ time, similarity search and classification over very long time series become practical where DTW's quadratic cost is prohibitive.","The single parameter $E$ controls noise sensitivity: excluding small $\\varepsilon$ values suppresses high-frequency noise, which the paper demonstrates can lift accuracy from 45% to 63% on the Worms dataset.","If the metric property survives, MDD can be used in metric-space indexing structures and algorithms that require the triangle inequality, not just in brute-force 1-NN.","MDD's envelope overlap avoids the pathological point-to-point alignments that plague DTW, because it never aligns individual time steps."],"supporting_citations":[{"why":"Supplies Dubuc's variation-envelope method that defines the multiscale envelopes used by MDD.","marker":"[10]"},{"why":"Establishes the triangle inequality for Jaccard distance, which the paper relies on to claim MDD is a metric.","marker":"[17]"},{"why":"Provides an earlier source for the distance-between-sets result, cited alongside [17] for the Jaccard triangle inequality.","marker":"[14]"},{"why":"Immediate predecessor and comparison method, also inspiring the multiscale intersection-over-union idea.","marker":"[2]"},{"why":"Supplies the 95 UCR datasets and the tuned-DTW baseline accuracies used in the performance comparison.","marker":"[9]"},{"why":"Provides the accuracy-gain evaluation framework (Texas Sharpshooter plot) used to quantify MDD's performance relative to baselines.","marker":"[4]"},{"why":"Defines the standard DTW baseline that MDD is compared against.","marker":"[20]"}],"fun_headline_variants":["MDD: a linear-time metric that rivals tuned DTW","Fractal+IoU metric matches DTW's accuracy on 95 datasets","Dubuc variation + Jaccard index: new time-series metric","MDD: fractal-based similarity measure proves to be a metric","New metric MDD matches tuned DTW, runs linear time"],"cache_read_input_tokens":12928,"weakest_assumption_plain":"The metric proof treats the total area under the ratio curve as if it were a single Jaccard distance, which only works if the scale weights sum to 1; with the paper's default scales they sum to 63, so the distance from a series to itself would be negative.","fun_headline_variants_meta":{"raw":{"variants":["MDD: a linear-time metric that rivals tuned DTW","Fractal+IoU metric matches DTW's accuracy on 95 datasets","Dubuc variation + Jaccard index: new time-series metric","MDD: fractal-based similarity measure proves to be a metric","New metric MDD matches tuned DTW, runs linear time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000682,"raw_usage":{"total_tokens":3156,"prompt_tokens":1065,"completion_tokens":2091,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":681,"completion_tokens_details":{"reasoning_tokens":2000}},"tokens_in":681,"tokens_out":2091,"duration_ms":15918,"temperature":1.0,"reasoning_tokens":2000,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:38:09.719782+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute Eq. 5 for a series paired with itself under the paper's generic scale set; with $E=\\{1,2,4,\\ldots,64\\}$ the area under the curve is 63, so $\\mathrm{MDD}(x,x)=1-63=-62$, contradicting the reflexivity axiom that the metric proof claims.","supporting_citations":[{"cited_title":"Physical review","cited_arxiv_id":null,"evidence_quote":"Supplies Dubuc's variation-envelope method that defines the multiscale envelopes used by MDD."},{"cited_title":"Pattern Recognit","cited_arxiv_id":null,"evidence_quote":"Establishes the triangle inequality for Jaccard distance, which the paper relies on to claim MDD is a metric."},{"cited_title":"Nature 239, 174–174 (1972)","cited_arxiv_id":null,"evidence_quote":"Provides an earlier source for the distance-between-sets result, cited alongside [17] for the Jaccard triangle inequality."},{"cited_title":"In: Amini, M.R., Canu, S., Fischer, A., Guns, T., Kralj Novak, P., Tsoumakas, G","cited_arxiv_id":null,"evidence_quote":"Immediate predecessor and comparison method, also inspiring the multiscale intersection-over-union idea."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the 95 UCR datasets and the tuned-DTW baseline accuracies used in the performance comparison."},{"cited_title":"In: Proceedings of the Eleventh SIAM International Conference on Data Mining, SDM 2011, April 28-30, 2011, Mesa, Arizona, USA","cited_arxiv_id":null,"evidence_quote":"Provides the accuracy-gain evaluation framework (Texas Sharpshooter plot) used to quantify MDD's performance relative to baselines."}],"review_version":1}