{"id":"0120868c-84a6-4de8-834a-88df32990c3a","arxiv_id":"1908.01794","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A review that restates known sufficient conditions for consistent clustering of ergodic stochastic processes and summarizes simulations from prior papers.","lead":"This paper reviews ways to group random sequences, called stochastic processes, into clusters, and concludes that consistent clustering follows from a dissimilarity measure that satisfies the triangle inequality together with a consistent estimator. It is a summary of earlier work, with no new theorems or data, and it contains errors in the covariance-based measure.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Type (2) dissimilarity is internally inconsistent: Definition 3.2 uses centered covariances and ignores means, while Definition 3.3 is uncentered, so hat d is not consistent for d and d does not match the stated ground truth.","rationale":"The reader's weakest assumption correctly identifies the uncentered estimator in Definition 3.3 as a consistency failure when means are nonzero; I agree this is a load-bearing flaw. My stress-test adds the equally serious theoretical mismatch that Definition 3.2's d is a centered-covariance-only dissimilarity, so it cannot characterize the Type (2) ground truth that includes equality of means. In fact, d assigns distance 0 to processes with different means but identical covariance, which are different clusters by the paper's own ground truth. These two problems together mean Theorem 2.1, as instantiated in Section 3, has false hypotheses: hat d is not consistent for d, and d does not separate the ground-truth classes. Since the paper is a review whose central contribution is the abstract's general conclusion, this unsupported Type (2) case is enough to reject. The simulation study in Section 5 cannot repair the inconsistency because it only reports decreasing mis-clustering rates for particular fBm and mBm examples and does not compare hat d with d. I therefore keep the reader's REJECT verdict unchanged.","tokens_in":6405,"tokens_out":6586,"duration_ms":64248,"concrete_test":"Simulate two iid Gaussian processes: X_t ~ N(0, 1) and Y_t ~ N(1, 1), each of length n, and set m_n = floor(log n). Compute the theoretical dissimilarity d from Definition 3.2: it is 0 because the centered covariance matrices of the two processes are identical for every block size m. Compute the sample dissimilarity hat d from Definition 3.3 on the simulated paths: for each m, the uncentered block matrix converges to (1 + mu_j^2) I_m, so hat d converges to a positive limit proportional to |mu_X^2 - mu_Y^2| * sum_m w_m sqrt(m), not to 0. Running this for n = 2^8, 2^12, 2^16 confirms that hat d does not decrease to 0, directly falsifying the consistency hypothesis needed by Theorem 2.1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's conclusion requires that for Type (2) the pair (d, hat d) satisfies the consistency hypothesis of Theorem 2.1 and that d characterizes the clustering ground truth. Both fail. In Definition 3.2, d(X^(1), X^(2)) is built from centered block covariance matrices Cov(X^(j)_l, ..., X^(j)_{l+m-1}) and therefore takes value 0 whenever two processes have identical covariance structure, regardless of their means. But the Type (2) ground truth defined in the introduction clusters processes with the same means and covariance structures, so two Gaussian processes with different means and the same covariance are in distinct ground-truth clusters while their theoretical dissimilarity is 0. Thus d does not metrize the stated equivalence relation. The estimator in Definition 3.3 is also incompatible: nu(x, l, m) = (1 / (n - m - l + 2)) * sum_{i=l}^{n-m+1} (X_i ... X_{i+m-1})^T (X_i ... X_{i+m-1}) is an uncentered sample second-moment matrix. For a stationary ergodic process with mean mu, it converges to E[X X^T] = Cov(X) + mu mu^T, not to Cov(X). Consequently hat d converges to a dissimilarity between raw second moments, not to d. The consistency assertion in Theorem 2.1 therefore has no valid instantiation in Section 3; the proof is delegated entirely to Khaleghi et al. (2016) and Peng et al. (2019), so these internal mismatches are never checked in this manuscript. The Type (3) construction inherits the Type (2) estimator, so the gap propagates to that section as well.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript surveys clustering algorithms for three classes of stochastic processes: distribution stationary ergodic, covariance stationary ergodic, and locally asymptotically self-similar. It presents two algorithms (offline and online) and Theorem 2.1, which claims asymptotic consistency when the number of clusters is known, the estimated dissimilarity is consistent, and both estimated and theoretical dissimilarity satisfy the triangle inequality. Section 3 defines a covariance-based dissimilarity for Type (2) processes, Section 4 adapts it to locally asymptotically self-similar processes, and Section 5 summarizes simulation experiments from the cited papers. The paper's central conclusion is that ergodicity plus the triangle inequality is a sufficient condition for consistent clustering.","tokens_in":6751,"tokens_out":5864,"duration_ms":60012,"significance":"If the claimed results held, the paper would offer a useful organizing principle for clustering stochastic processes and would tie together recent algorithms for covariance-stationary and locally asymptotically self-similar processes. The explicit statement of the sufficient condition is a helpful expository contribution, and the comparative simulation summary adds context. However, the positive value is limited by the absence of new proofs and by a load-bearing mathematical inconsistency in the Section 3 dissimilarity: the theoretical and empirical dissimilarities do not match, and the theoretical dissimilarity does not characterize the stated ground truth. No machine-checked proofs or reproducible code are provided, and the main theorem is proved entirely by citation to prior work.","major_comments":[{"comment":"The theoretical dissimilarity d does not metrize the stated ground truth for Type (2). The ground truth defined in the Introduction clusters processes with the same means and covariance structures. Definition 3.2, however, is built from centered covariance matrices Cov(X^(1)_l, ..., X^(1)_{l+m-1}) and contains no mean term. For two stationary Gaussian processes with different means and identical covariance matrices, the ground truth places them in different clusters, while d equals 0. Hence d does not characterize the Type (2) equivalence relation, and Theorem 2.1 cannot be instantiated with this d.","section":"Introduction and Definition 3.2"},{"comment":"The estimator d-hat is not consistent for the d of Definition 3.2. The quantity ν(x, l, m) is defined as the average of outer products (X_i ... X_{i+m-1})^T (X_i ... X_{i+m-1}) without subtracting the mean. For a covariance stationary ergodic process with mean μ, ν converges to E[X X^T] = Cov(X) + μ μ^T rather than to Cov(X). Unless every process under comparison has mean zero, d-hat converges to a dissimilarity between raw second moments, not to the centered dissimilarity d. The consistency hypothesis of Theorem 2.1 is therefore not satisfied by the Section 3 example.","section":"Definition 3.3"},{"comment":"The proof of Theorem 2.1 is entirely delegated to Khaleghi et al. (2016) and Peng et al. (2019), and the manuscript never verifies that the pair (d, d-hat) defined in Section 3 satisfies the stated hypotheses. Because of the mismatches noted above, this verification is not merely omitted but false. The Type (3) estimator in Eq. (4.4) is built from the same uncentered estimator of Definition 3.3, so the problem propagates to Theorem 4.2 as well.","section":"Theorem 2.1 and Section 4"}],"minor_comments":[{"comment":"Figure 1 is referenced and described in the caption but is not included in the manuscript, so the simulation claims cannot be checked.","section":"Section 5 and Figure 1"},{"comment":"The reference \"et al., M. (2018)\" is incomplete and unverifiable; the author names and full citation information need to be supplied.","section":"References"},{"comment":"The \"log*-transformation\" mentioned in the simulation summary for fractional Brownian motion is not defined anywhere in the paper.","section":"Section 5.1"},{"comment":"The notation Cov(X^(j)_l, ..., X^(j)_{l+m-1}) should be explicitly defined as the covariance matrix of the indicated block, and the normalization or summability condition on the weights w_j should be stated precisely.","section":"Definition 3.2"},{"comment":"There are several typographical errors and infelicities, including \"til 2015\" in the Introduction, \"Sceinces\" in the affiliation, \"in proved\" in the proof of Theorem 2.1, and \"khaleghi\" at the start of Section 5; a careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"From the editor's perspective, the manuscript is largely a summary of the authors' own prior work (Peng et al. 2018, 2019), and the main theorem is proved by citation to those works. The internal inconsistency in Section 3 is not a presentation issue; it invalidates the claimed example and therefore the central message. A revision would require redefining the dissimilarity to include mean differences and supplying a valid consistency proof for the new estimator, which goes beyond local editing and would amount to new research. I do not see a path to acceptance within the current manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a review, not a new contribution. It pulls together work from Khaleghi et al. (2016) and Peng et al. (2018, 2019) and states a clean sufficient condition for clustering consistency: if the theoretical dissimilarity and its estimator both satisfy the triangle inequality and the estimator is consistent, then the two algorithms are consistent. As a summary of that line of work, the high-level idea is fine.\n\nThe review does what it says: it explicitly says 'we review' and gives simulation summaries from the original papers. The examples chosen are representative. If you want a one-page entry point to this literature, the framing is useful.\n\nThe problems are real, though. The Type (2) section does not hang together. Definition 3.2 defines d via centered covariance matrices, while Definition 3.3 defines \\hat d with uncentered sample second moments. For a stationary ergodic process with mean \\mu, the estimator converges to E[XX^T] = Cov(X) + \\mu\\mu^T, not Cov(X). So \\hat d is not consistent for d. The stress-test note is correct on this. On top of that, the Type (2) ground truth requires equal means, but d only compares covariances, so two Gaussian processes with the same covariance and different means are in different ground-truth clusters while d = 0. The dissimilarity does not metrize the stated equivalence relation. The abstract's claim that the Type (2) example instantiates the sufficient condition is therefore false as written.\n\nThe theorems themselves are not proved here. That is fine for a review, but it means the internal mismatch between d and \\hat d was never caught, and the reader is sent back to the original papers. The centered/uncentered issue is not a research-level subtlety; it is a basic property of second moments.\n\nWho is this for? Someone who wants a quick bibliography and the general theorem statement. But I would not rely on it for the Type (2) details. I would not send it to peer review in its current form; the corrections are substantial enough that the review needs rewriting. If the authors fix the estimator to center the data and adjust the dissimilarity or the ground truth, it could become a passable survey.","headline":"A clearly-written review of a niche clustering literature whose Type (2) example is internally inconsistent: the estimator is uncentered while the target dissimilarity is centered.","tokens_in":7274,"tokens_out":2792,"would_cite":false,"duration_ms":28528,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30","62M10","60G10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Ergodic stochastic processes can be clustered consistently when the dissimilarity measure and its estimate both satisfy the triangle inequality.","keywords":["stochastic process clustering","ergodic processes","dissimilarity measure","triangle inequality","asymptotic consistency","covariance stationary ergodic","local asymptotic self-similarity","unsupervised learning"],"falsifier":"Sample two long paths from a covariance-stationary ergodic process with a nonzero mean, compute the empirical dissimilarity of Definition 3.3 for increasing $n$ with $m_n=o(n)$, and compare it with the theoretical $d$ of Definition 3.2; if the estimates do not converge to $d$ because the product terms are not centered, the consistency premise fails. Equivalently, compute the estimator on a constant-shifted version of a zero-mean process and check whether it changes: a dissimilarity targeting covariances should be invariant to additive shifts, while the uncentered estimator will not be.","tokens_in":6173,"feed_emoji":"📊","tokens_out":9608,"duration_ms":86394,"temperature":0.7,"pith_summary":"This paper is a review that extracts a single sufficient condition for clustering stochastic processes consistently: the processes must be ergodic in the relevant sense, and the dissimilarity measure used by the algorithm, together with its sample estimate, must satisfy the triangle inequality and converge to the population dissimilarity as sample paths grow. Under that condition, the same offline and online clustering algorithms are asymptotically consistent for distribution-stationary ergodic processes, covariance-stationary ergodic processes, and, in an approximate sense, locally asymptotically self-similar processes. The paper presents this as the common structure behind earlier consistency proofs. It illustrates the claim with fractional Brownian motion and multifractional Brownian motion simulations, where mis-clustering rates fall as path length grows.","feed_headline":"Ergodic processes cluster consistently under triangle inequality","feed_subtitle":"One condition covers distribution-ergodic, covariance-ergodic, and self-similar process data.","key_machinery":"The load-bearing object is the dissimilarity pair $(d,\\hat d)$: a population dissimilarity $d$ between two processes and a sample dissimilarity $\\hat d$ computed from finite paths, with both obeying the triangle inequality and $\\hat d$ converging to $d$ as the shorter path length grows. For covariance-stationary ergodic processes, $d$ is a weighted, doubly infinite sum over lags of the Frobenius-norm distance between $m\\times m$ covariance matrices of the two processes, and $\\hat d$ replaces population covariances with overlapping-block sample covariances using block size $m_n=o(n)$. The triangle inequality is what lets the farthest-point seeding and nearest-center assignment steps transmit estimation error into cluster boundaries; the consistency of $\\hat d$ is what makes the error vanish. For locally asymptotically self-similar processes, the extra machinery is a localized increment transformation: increments over small windows behave approximately like increments of a stationary ergodic tangent process, so the same dissimilarity can be applied locally.","core_discovery":"The central claim, stated as Theorem 2.1, is that Algorithms 1 and 2, an offline farthest-distance seeding procedure and an online weighted version of it, are asymptotically consistent for clustering distribution-stationary ergodic and covariance-stationary ergodic processes whenever the number of clusters is known, the sample dissimilarity $\\hat d$ is a consistent estimator of the theoretical dissimilarity $d$, and both $d$ and $\\hat d$ satisfy the triangle inequality. For locally asymptotically self-similar processes, the same conclusion holds in an approximate sense: local increment paths are approximately distributed as covariance-stationary ergodic increment processes, so applying the covariance dissimilarity to localized increments gives an approximately consistent clustering. The authors frame this as a sufficient condition, not a characterization, and present it as the common structure behind earlier consistency proofs.","pith_inferences":["A centered variant of the covariance-block estimator would be a natural direct test of the theory: the estimator as written does not subtract the mean, so it should converge to the covariance dissimilarity for zero-mean processes and to a second-moment dissimilarity for shifted processes.","The theorem's conditions are sufficient but not necessary, so one can test whether violating the triangle inequality destroys consistency even when the dissimilarity estimator is perfectly consistent.","The localized-increment construction suggests a generic recipe for nonstationary processes: first transform into stationary ergodic increments, then apply a triangle-inequality dissimilarity on the transformed paths; this could be checked on other locally stationary models."],"forward_implications":["Offline and online clustering of distribution-stationary ergodic and covariance-stationary ergodic processes is asymptotically consistent when the true number of clusters is known and the dissimilarity condition holds.","The same consistency guarantee transfers to any process class for which a consistent triangle-inequality dissimilarity can be constructed, because the theorem separates the probabilistic assumption from the algorithm.","For locally asymptotically self-similar processes, clustering is approximately asymptotically consistent, with the approximation governed by the local self-similarity convergence in Eq. (4.1).","In the simulation examples, mis-clustering rates for fractional Brownian motion and multifractional Brownian motion decrease as path length grows, matching the theoretical consistency statement."],"supporting_citations":[{"why":"Proves the consistency of Algorithms 1 and 2 for distribution-stationary ergodic processes, the baseline result Theorem 2.1 extends.","marker":"Khaleghi et al., 2016"},{"why":"Supplies the covariance-based dissimilarity pair (Definitions 3.2 and 3.3) and proves consistency for covariance-stationary ergodic processes.","marker":"Peng et al., 2019"},{"why":"Establishes the approximate consistency result for locally asymptotically self-similar processes using localized increment paths.","marker":"Peng et al., 2018"},{"why":"Provides the definition and path properties of locally asymptotically self-similar processes used in Assumption (A).","marker":"Boufoussi et al., 2008"}],"fun_headline_variants":["Clustering ergodic processes: triangle inequality suffices","Consistent clustering for ergodic processes under triangle inequality","Triangle inequality yields consistent clustering of ergodic processes","Ergodic process clustering consistent under triangle inequality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument leans on the assumption that the dissimilarity computed from finite sample paths converges to the true dissimilarity for every covariance-stationary ergodic process, with the block size growing more slowly than the sample length; that convergence is cited from earlier work rather than proved here.","fun_headline_variants_meta":{"raw":{"variants":["Clustering ergodic processes: triangle inequality suffices","Consistent clustering for ergodic processes under triangle inequality","Triangle inequality yields consistent clustering of ergodic processes","Ergodic process clustering consistent under triangle inequality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000522,"raw_usage":{"total_tokens":2422,"prompt_tokens":736,"completion_tokens":1686,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":352,"completion_tokens_details":{"reasoning_tokens":1625}},"tokens_in":352,"tokens_out":1686,"duration_ms":12803,"temperature":1.0,"reasoning_tokens":1625,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:03:15.208435+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Sample two long paths from a covariance-stationary ergodic process with a nonzero mean, compute the empirical dissimilarity of Definition 3.3 for increasing $n$ with $m_n=o(n)$, and compare it with the theoretical $d$ of Definition 3.2; if the estimates do not converge to $d$ because the product terms are not centered, the consistency premise fails. Equivalently, compute the estimator on a constant-shifted version of a zero-mean process and check whether it changes: a dissimilarity targeting covariances should be invariant to additive shifts, while the uncentered estimator will not be.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Proves the consistency of Algorithms 1 and 2 for distribution-stationary ergodic processes, the baseline result Theorem 2.1 extends."},{"cited_title":"Covariance-based Dissimilarity Measures Applied to Clustering Wide-sense Stationary Ergodic Processes","cited_arxiv_id":"1801.09049","evidence_quote":"Supplies the covariance-based dissimilarity pair (Definitions 3.2 and 3.3) and proves consistency for covariance-stationary ergodic processes."},{"cited_title":"Cluster Analysis on Locally Asymptotically Self-similar Processes with Known Number of Clusters","cited_arxiv_id":"1804.06234","evidence_quote":"Establishes the approximate consistency result for locally asymptotically self-similar processes using localized increment paths."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the definition and path properties of locally asymptotically self-similar processes used in Assumption (A)."}],"review_version":1}