Pith. sign in

REVIEW 3 major objections 5 minor 23 references

Some Developments in Clustering Analysis on Stochastic Processes

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Ergodic stochastic processes can be clustered consistently when the dissimilarity measure and its estimate both satisfy the triangle inequality.

desk verdict 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. read the letter →

arxiv 1908.01794 v1 pith:7WDUTCRX submitted 2019-08-05 stat.ML cs.LG

Qidi Peng , Nan Rao , Ran Zhao This is my paper · ORCID
classification stat.MLcs.LG MSC 62H3062M1060G10
keywords stochasticprocessclusteringergodicprocessesdissimilaritymeasuretriangleinequalityasymptoticconsistencycovariancestationarylocalself-similarityunsupervisedlearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

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.

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 (3)
  1. [Introduction and Definition 3.2] 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.
  2. [Definition 3.3] 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.
  3. [Theorem 2.1 and Section 4] 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.
minor comments (5)
  1. [Section 5 and Figure 1] Figure 1 is referenced and described in the caption but is not included in the manuscript, so the simulation claims cannot be checked.
  2. [References] The reference "et al., M. (2018)" is incomplete and unverifiable; the author names and full citation information need to be supplied.
  3. [Section 5.1] The "log*-transformation" mentioned in the simulation summary for fractional Brownian motion is not defined anywhere in the paper.
  4. [Definition 3.2] 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.
  5. [Throughout] 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.

Circularity Check

3 steps flagged · score 5.0 of 10

The central Type (2) and Type (3) consistency claims are imported from the authors' own prior papers (Peng et al. 2019, 2018), and the printed Type (2) estimator is uncentered, so the claimed consistency instantiation is not established.

  1. self citation load bearing [Section 2, Theorem 2.1 and its proof]
    "Theorem 2.1 Algorithms 1 and 2 are asymptotically consistent for the processes of Types (1) and (2) respectively, provided that the correct number κ of clusters is known, and the sample dissimilarity measure ˆd is consistent and both ˆd and d satisfy the triangle inequality. Proof. The consistency of Algorithms 1 and 2 applied for clustering processes of Type (1) is proved in (Khaleghi et al., 2016); the consistency of the two algorithms applied for clustering processes of Type (2) in proved in (Peng et al., 2019)."

    The central Type (2) result is not derived in this manuscript. The proof is a citation to Peng et al. (2019), whose authors are the same as the present paper. The theorem is therefore load-bearing on a self-citation, and the paper supplies no independent check that the specific d and ˆd of Section 3 satisfy the theorem's consistency and triangle-inequality hypotheses.

  2. other [Section 3, Definitions 3.2 and 3.3]
    "ν(x,l,m) := \sum_{i=l}^{n-m+1} (X_i...X_{i+m-1})^T (X_i...X_{i+m-1}) / (n-m-l+2) are the estimators of stationary covariance matrices."

    The estimator is an uncentered outer-product average, while Definition 3.2 defines d via centered Cov(...) matrices. For a stationary ergodic process with mean μ, ν converges to E[XX^T] = Cov(X) + μμ^T, not Cov(X). Also, d ignores means while the Type (2) ground truth from the Introduction clusters processes with the same means and covariance structures. Thus the printed (d, ˆd) pair does not satisfy Theorem 2.1's consistency condition; the claimed Type (2) consistency is not established for the definitions as written. This is a correctness gap rather than a definitional equivalence, but it makes the self-cited theorem unverified in this text.

1 more flagged steps
  1. self citation load bearing [Section 4, Theorem 4.2]
    "Accordingly, the consistency of Algorithms 1 and 2 can be expressed in the following way: Theorem 4.2 Under Assumption (A), Algorithms 1 and 2 are approximately asymptotically consistent, if ˆd is replaced with ˆd∗."

    The statement is presented as following from the approximate f.d.d. equality (4.2), which is itself attributed to 'Peng et al. (Peng et al., 2018) showed...' — a same-author paper. No proof is given here. The Type (3) consistency claim is therefore inherited from the authors' prior work rather than established in this manuscript.

full rationale

This manuscript is explicitly a review ('We review some developments'), so most of its content is, by design, a survey of results proved elsewhere. That structure by itself is not circular. However, the main theorem of the paper, Theorem 2.1, is stated as a theorem of this manuscript, and its proof consists only of citations: for Type (1) to Khaleghi et al. (2016), and for Type (2) to Peng et al. (2019), which shares all three present authors. The Type (3) consistency claim in Theorem 4.2 is similarly derived from Peng et al. (2018), also by the same authors. These are load-bearing self-citations for the paper's central conclusion, although they are not algebraic reductions of the result to its own statement. Independently, Section 3 contains a concrete mismatch: Definition 3.2 defines d through centered covariance matrices, while Definition 3.3's ν is an uncentered sample second-moment estimator; for a process with nonzero mean the estimator does not converge to the centered covariance, so the pair (d, ˆd) as printed does not satisfy the consistency hypothesis of Theorem 2.1. This is a correctness gap rather than a circularity, but it means the self-cited theorem is not verified by the definitions printed here. Taking these together, the central claim is not derived in this paper; it is inherited from same-author sources, with an additional unverified estimator consistency. The score is therefore moderate, not extreme.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper's central claim depends on the dissimilarity estimator being consistent and on the triangle inequality. These are assumed or delegated to prior works; no new proof, code, or data are supplied.

free parameters (4)
  • weight sequence {w_j}
    Positive weights chosen so that the theoretical dissimilarity d is finite; the paper gives no specific choice, and clustering outputs depend on them.
  • block size m_n
    Chosen to be o(n) in Definition 3.3; the rate is unspecified, and uniform consistency of the estimator over lags up to m_n is assumed.
  • K in local increment paths
    Arbitrary positive integer in Eq. (4.2) used to define localized increment paths in Eq. (4.3); the approximation quality depends on it.
  • time scale tau
    Equation (4.2) requires tau to be sufficiently small for the tangent-process approximation, but no rate or error bound is provided.
assumptions (4)
  • domain assumption The true and empirical dissimilarity measures d and d-hat both satisfy the triangle inequality.
    Stated as a core condition for Theorem 2.1. For the covariance-based d this holds because a weighted sum of Frobenius norms satisfies the triangle inequality, but for d-hat this must also hold at every finite n.
  • domain assumption d-hat is a consistent estimator of d.
    Required for Theorem 2.1. For the covariance-based estimator this is not proved in the paper and is delegated to Peng et al. (2019).
  • domain assumption Processes are distribution stationary ergodic or covariance stationary ergodic.
    Definition 3.1 and the surrounding text restrict the class of processes; consistency is only claimed under these assumptions.
  • domain assumption Assumption (A) for locally asymptotically self-similar processes.
    Requires distinct functional indexes H and covariance stationary ergodic tangent increment processes; used for Theorem 4.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Some Developments in Clustering Analysis on Stochastic Processes." pith.science (2026). https://pith.science/paper/7WDUTCRX

@misc{pith2026190801794,
  author       = {Pith},
  title        = {Pith review of: Some Developments in Clustering Analysis on Stochastic Processes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7WDUTCRX}},
  note         = {Machine review of arXiv:1908.01794}
}
read the original abstract

We review some developments on clustering stochastic processes and come with the conclusion that asymptotically consistent clustering algorithms can be obtained when the processes are ergodic and the dissimilarity measure satisfies the triangle inequality. Examples are provided when the processes are distribution ergodic, covariance ergodic and locally asymptotically self-similar, respectively.

Figures

Figures reproduced from arXiv: 1908.01794 by the authors.

Figure 1
Figure 1. The top graph illustrates the mis-clustering rates of offline algorithm [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages

  1. [1]

    S., and Wah, T

    Aghabozorgi, S., Shirkhorshidi, A. S., and Wah, T. Y. (2015). Time-series clustering - a decade review. Information Systems , 53:16--38

    work page 2015

  2. [2]

    Ayache, A., Cohen, S., and L \'e vy-V \'e hel, J. (2000). The covariance structure of multifractional B rownian motion, with application to long range dependence. In Acoustics, Speech, and Signal Processing, 2000. ICASSP'00. Proceedings. 2000 IEEE International Conference on , volume 6, pages 3810--3813. IEEE

    work page 2000

  3. [3]

    and L \'e vy-V \'e hel, J

    Ayache, A. and L \'e vy-V \'e hel, J. (2004). On the identification of the pointwise H \"o lder exponent of the generalized multifractional B rownian motion. Stochastic Processes and their Applications , 111(1):119--156

    work page 2004

  4. [4]

    Boufoussi, B., Dozzi, M., and Guerbaz, R. (2008). Path properties of a class of locally asymptotically self-similar processes. Electronic Journal of Probability , 13(29):898--921

    work page 2008

  5. [5]

    Buard, B., Humeau, A., Rousseau, D., Chapeau-Blondeau, F., and Abrahamc, P. (2010). Pointwise H \"older exponents of a model for skin laser doppler flowmetry signals based on six nonlinear coupled oscillators with linear and parametric couplings: Comparison with experimental data from young healthy subjects. IRBM , 31:175--181

    work page 2010

  6. [6]

    Damian, D., Ore s i c , M., and Verheij, E. e. a. (2007). Applications of a new subspace clustering algorithm ( COSA ) in medical systems biology. Metabolomics , 3(1):69--77

    work page 2007

  7. [7]

    et al., M. (2018). Clustering gene expression time series data using an infinite G aussian process mixture model. PLoS Comput Biol , 14(1):e1005896

    work page 2018

  8. [8]

    Falconer, K. (2002). Tangent fields and the local structure of random fields. Journal of Theoretical Probability , 15(3):731--750

    work page 2002

Show all 23 references
  1. [9]

    and Franosch, T

    H\"ofling, F. and Franosch, T. (2013). Anomalous transport in the crowded world of biological cells. Reports on Progress in Physics , 76(4):046602

    2013

  2. [10]

    Humeau, A., Chapeau-Blondeau, F., Rousseau, D., Tartas, M., Fromy, B., and Abraham, P. (2007). Multifractality in the peripheral cardiovascular system from pointwise H \"older exponents of laser doppler flowmetry signals. iophysical Journal , 93(12):L59--L61

    2007

  3. [11]

    J\"a\"askinen, V., Parkkinen, V., Cheng, L., and Corander, J. (2014). Bayesian clustering of DNA sequences using markov chains and a stochastic partition model. Stat. Appl. Genet. Mol. Biol. , 13(1):105--121

    2014

  4. [12]

    and Ryabko, D

    Khaleghi, A. and Ryabko, D. (2012). Locating changes in highly dependent data with unknown number of change points. In Advances in Neural Information Processing Systems , pages 3086--3094

    2012

  5. [13]

    and Ryabko, D

    Khaleghi, A. and Ryabko, D. (2014). Asymptotically consistent estimation of the number of change points in highly dependent time series. In International Conference on Machine Learning , pages 539--547

    2014

  6. [14]

    Khaleghi, A., Ryabko, D., Mari, J., and Preux, P. (2016). Consistent algorithms for clustering time series. Journal of Machine Learning Research , 17(3):1--32

    2016

  7. [15]

    and van Ness, J

    Mandelbrot, B. and van Ness, J. W. (1968). Fractional B rownian motions, fractional noises and applications. SIAM Review , 10(4):422--437

    1968

  8. [16]

    T., Leier, A., and Burrage, K

    Marquez-Lago, T. T., Leier, A., and Burrage, K. (2012). Anomalous diffusion and multifractional brownian motion: simulating molecular crowding and physical obstacles in systems biology. IET Systems Biology , 6(4):134--142

    2012

  9. [17]

    Maruyama, G. (1970). Infinitely divisible processes. Theory of Probability and its Applications , 15(1):1--22

    1970

  10. [18]

    and V \'e hel, J

    Peltier, R.-F. and V \'e hel, J. L. (1995). Multifractional B rownian motion: definition and preliminary results. apport de recherche de l'INRIA, INRIA , no. 2645:239--265

    1995

  11. [19]

    Peng, Q., Rao, N., and Zhao, R. (2018). Clustering analysis on locally asymptotically self-similar processes with known number of clusters. arXiv , 1804.06234v1

    2018 arXiv

  12. [20]

    Peng, Q., Rao, N., and Zhao, R. (2019). Covariance-based dissimilarity measures applied to clustering wide-sense stationary ergodic processes. Accepted by Machine Learning , arXiv, 1801.09049v4

    2019 arXiv

  13. [21]

    \`Slezak, J. (2017). Asymptotic behaviour of time averages for non-ergodic G aussian processes. Annals of Physics , 383:285--311

    2017

  14. [22]

    Taqqu, M. S. (2013). Beno\^it M andelbrot and fractional B rownian motion. Statistical Science , 28(1):131--134

    2013

  15. [23]

    Zhao, W., Zou, W., and Chen, J. J. (2014). Topic modeling for cluster analysis of large biological and medical datasets. BMC Bioinformatics , 15:S11

    2014

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.