REVIEW 4 major objections 5 minor 13 references
Modelling Loss of Complexity in Intermittent Time Series and its Application
T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash
Pith's one-line read The paper claims that a nonparametric relative-entropy summary, with BIC-selected lag and CUSUM detection, locates complexity change-points in intermittent time series more accurately than approximate entropy.
desk verdict A real methodological extension of relative entropy to m-consecutive variables with a Gaussian limit, but the empirical claim that it beats ApEn is not yet established because the main comparison is confounded. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the nonparametric relative entropy $\hat E(m,h|X)$ for $m$ consecutive variables, a Kullback-Leibler divergence between the joint density of $m+1$ consecutive observations and the product of the $m$-history density and the marginal density. The paper estimates all three densities with leave-one-out Jackknife kernel estimators, selects the lag $m$ by BIC on a Nadaraya-Watson nonlinear autoregression, and then feeds the resulting scalar sequence into a CUSUM change-point detector. This scalar map is what turns each intermittent time series into one comparable number, so that breakpoints among segments become breakpoints in the mean of the RlEn sequence.
What would settle it
Take two stationary AR(3) processes with a known change-point, generate intermittent series of length 500, then multiply the raw series by constants 0.1, 1, and 10 before the logistic transform and RlEn computation; if the detected change-point or the RlEn values shift by more than simulation error, the transformation-invariance claim fails.
Extended reading notes
Core claim
The central claim is that mapping each intermittent time series to its nonparametric relative entropy $E = E_f[\log(f(X_{i;m+1})/(g(X_{i;m})g_1(X_{i+m})))]$, with the lag $m$ chosen by a BIC criterion, produces a scalar sequence whose CUSUM change-points coincide with the times at which the underlying process loses complexity, and do so more accurately than approximate entropy. The estimator uses leave-one-out Jackknife kernel density estimates; under strict stationarity and a marginal density bounded away from zero on $[0,1]$, the paper proves a normal limiting distribution at rate $\sqrt{n h^{(m+1)/2}}$ and consistency of the BIC lag selector. The method is designed to be transformation invariant and background noise-free, properties that mean, variance, entropy, and conditional entropy summaries do not have. In the muscle-fatigue application, the detected change-point separates contractions into two groups whose mean RlEn differs highly significantly.
Load-bearing premise
The method's proof assumes each time series is strictly stationary and its distribution never touches zero on the unit interval, but the simulations and the real muscle data include non-stationary series, so the theoretical guarantee does not cover the conditions under which the method is actually used.
Editorial extensions
If this is right
- Because RlEn is transformation invariant and background noise-free, change-point detections should not move when the measurement unit or noise variance of the intermittent signals changes.
- With $m$ selected by BIC, RlEn keeps a small mean absolute distance from the true change-point even when $m$ is misspecified, whereas ApEn's accuracy collapses for larger $m$ in the AR(3) simulation.
- The method detects a statistically reliable fatigue change-point in all 11 subjects of the muscle dataset, while ApEn fails to detect one for Subject 4 and gives an unreliable point for Subject 1.
- The limiting Gaussian distribution of $\hat E(m)$ justifies using the RlEn sequence as input to standard CUSUM-type change-point procedures.
Reading between the lines
- If the invariance claim holds operationally, RlEn could serve as a unit-free feature for cross-subject or cross-device comparisons in clinical monitoring, where force, torque, or EEG scales differ across recordings.
- The paper's own numbers indicate a potential gap: after the logistic transform only 58.11% of the AR(3) series are stationary, yet RlEn still performs well; a natural test is whether the asymptotic normality and BIC consistency survive under non-stationarity or whether those simulations are too mild to expose the gap.
- One could extend the method to multivariate intermittent series by replacing the scalar RlEn with a vector of pairwise or block RlEn values, though the current CUSUM framework would not directly accept such a vector.
- Because RlEn is computed per segment, the two-stage detector propagates any BIC lag-selection error into the entropy values; the paper's robustness results suggest the propagation is mild, but a formal sensitivity analysis is not provided.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a nonparametric relative-entropy statistic (RlEn) as a scalar summary of each segment in a collection of intermittent time series, then applies a CUSUM-type change-point detector to the sequence of RlEn values. The lag order m is selected by BIC from a nonlinear autoregressive fit, and the entropy is estimated with jackknife kernel density estimators after a logistic transformation to [0,1]. The authors derive a Gaussian limit for the relative-entropy estimator under a fixed bandwidth and prove consistency of the BIC lag selection. They compare RlEn with approximate entropy (ApEn) in three simulation cases and in a real dataset of fatiguing muscle contractions, claiming that RlEn detects complexity change-points more accurately than ApEn.
Significance. If the comparison with ApEn were clean, the paper would fill a practical gap: ApEn is sensitive to two free parameters and is not transformation invariant, while RlEn has a nonparametric justification and comes with convergence-rate guarantees. The paper also ships reproducible code on GitHub, and the BIC and entropy components have independent foundations in Vieu (1995) and Hong and White (2005), so the method is not circularly constructed. However, the headline empirical claim is currently not established because the ApEn comparison in the key simulation is confounded and the ApEn tuning parameters are unreported; the theoretical results also do not cover the non-stationary operating conditions in which the method is applied.
major comments (4)
- [§4.2, Table 3] The comparison in Case 2 is not matched: after logistic transformation only 58.11% of the series are stationary, and the paper states that ApEn results are based entirely on the stationary subset while RlEn is applied to all 150 repetitions. Because the change-point at 61 refers to the original ordering, it does not necessarily correspond to position 61 in the stationary subsequence, and the ApEn input vector is roughly half as long. The large ApEn failure counts (e.g., 110/150 for m=2) are therefore confounded with sample-size reduction and index misalignment. This undermines the abstract's headline claim that RlEn outperforms ApEn.
- [§4.1, §4.2, §5, and Tables 2, 3, 5, 6] The ApEn tolerance parameter r is never reported: the description "follow the suggestions in Pincus (1991)" in §5 is not a unique specification, and r is known to strongly affect ApEn values. Without the r value, none of the ApEn comparisons in Tables 2, 3, 5, and 6 is reproducible or auditable. The authors should report r for every ApEn computation and ideally provide a sensitivity analysis over r.
- [§2 and Theorem 3.1] There is a gap between theory and implementation for the bandwidth: Theorem 3.1 gives a Gaussian limit for \hat E(m,h) under rate conditions on a fixed h, but in §2 and Algorithm 2 h is chosen data-dependently by maximizing (or, as written in Algorithm 2, minimizing) \hat E(\hat m,h|X_j). No proof is given that the CLT survives data-driven h, and no sensitivity analysis over h is reported. Additionally, Algorithm 2's "arg min_h" contradicts the text's "arg max_h".
- [Appendix B and §3, in relation to §4.2 and §5] The paper's invariance and noise-free selling points are broader than the results: Proposition B.1 establishes background-noise-freeness only for stationary Gaussian AR(p) processes, while the method is applied to general nonlinear and explicitly non-stationary series (41.89% of Case 2 series, and the real muscle data). Theorem 3.1 requires strict stationarity and a marginal density bounded away from zero on [0,1]; no result is provided for the non-stationary operating conditions, so the sentence "our method is robust even for non-stationary time series as well" (§4.2) is not supported by the theory.
minor comments (5)
- [Algorithm 2, line 4] The objective should be arg max_h, not arg min_h, to match the text immediately before it.
- [Theorem 3.1 and Theorems D.2–D.3] The condition written "nhm+13 → 0" is ambiguous; it should be typeset as n h^{m+13} or the intended power should be stated explicitly.
- [Table 2, row 'Mean'] The detected fraction for the mean method is 1/150, which is 0.67%, not 0.007% as printed.
- [§5, paragraph after Table 5] The sentence "Subject 2's is just 21.84%" appears to be a typo; Table 4 gives Subject 10 the 21.84% value.
- [Algorithm 1, line 10] The leave-one-out estimator is denoted \hat g^{-i} in the algorithm but is defined as \hat F^{-i} in equation (4); the notation should be unified.
Circularity Check
No material circularity: the RlEn estimator, BIC lag selection, and CUSUM change-point pipeline are independently grounded, and the self-citations are peripheral.
full rationale
The paper's derivation chain does not reduce to its own inputs. The RlEn estimator (Eq. 2) is a leave-one-out kernel plug-in of the KL divergence, with asymptotic normality (Theorem 3.1) proved by adapting Hong and White (2005) to m-consecutive variables; Hong and White is an external statistical reference, not a result of this paper. The BIC lag selection (Eq. 6) is justified via Vieu (1995) and Shao (1997), with conditions (C1)-(C7) stated and a proof of consistency in Theorem 2.2. Nothing in that chain uses the target change-point locations as an input. The transformation-invariance and background-noise-free claims are derived in Appendices A and B (e.g., Eq. 15 and Prop. B.1), not assumed. The change-points in Tables 2, 3, 6, and the real-data analysis are empirical outputs of the CUSUM/PELT algorithm applied to the RlEn scalar sequence, so they are not fitted parameters relabelled as predictions. Self-citations (e.g., Li et al. 2024 on deep-learning change-point detection and heating demand clustering) appear as related work and are not load-bearing for the proposed method. The main legitimate concern is validity of the Case 2 ApEn comparison, where the paper itself notes after logistic transformation only 58.11% of series are stationary and 'ApEn results are totally based on the stationary time series,' while RlEn uses all 150 repetitions; this asymmetry, and the unreported ApEn tolerance r, could confound the 'outperforms ApEn' claim. However, this is a statistical fairness/correctness issue, not circularity: it does not make RlEn's construction equivalent to the data being predicted. I therefore find no circular step and assign score 0.
Assumptions & free parameters
free parameters (3)
- Bandwidth h for relative entropy estimator =
data-dependent, selected by maximizing E(hat m, h|X_j)
- Bandwidth h* for nonlinear autoregression =
data-dependent, selected by leave-one-out cross-validation
- Maximum lag M =
10 (default)
assumptions (5)
- domain assumption The time series {X_t} is strictly stationary with support I=[0,1], marginal density G bounded away from zero, twice continuously differentiable with Lipschitz second derivative (Assumption 3.1).
- standard math The kernel K is symmetric with compact support [-1,1], integrates to 1, and has finite second moment (Assumption 2.1).
- domain assumption The series satisfies alpha-mixing with exponential decay and moment conditions (C1-C7 in Appendix C).
- domain assumption For the background noise-free claim, the process is assumed Gaussian AR(p) (Appendix B).
- ad hoc to paper The logistic transformation maps data to [0,1] and the estimator assumes the density is bounded away from zero on this support.
Cite this review
Pith. "Pith review of Modelling Loss of Complexity in Intermittent Time Series and its Application." pith.science (2026). https://pith.science/paper/EYXBND44
@misc{pith2026241114635,
author = {Pith},
title = {Pith review of: Modelling Loss of Complexity in Intermittent Time Series and its Application},
year = {2026},
howpublished = {\url{https://pith.science/paper/EYXBND44}},
note = {Machine review of arXiv:2411.14635}
}
read the original abstract
In this paper, we developed a novel method of nonparametric relative entropy (RlEn) for modelling loss of complexity in intermittent time series. The method consists of two steps. We first fit a nonlinear autoregressive model to each intermittent time series, where the corresponding lag order and the loss of complexity are determined by Bayesian Information Criterion (BIC) and relative entropy respectively. Then, change-points in the complexity are detected by a cumulative sum (CUSUM) based statistic. Compared to approximate entropy (ApEn), a popular method in literature, the performance of RlEn was assessed by simulations in terms of (1) ability to localize complexity change-points in intermittent time series; (2) ability to faithfully estimate underlying nonlinear models. The performance of the proposal was then examined in a real analysis of fatigue-induced changes in the complexity of human motor outputs. The results showed that the proposed method outperformed the ApEn in accurately detecting changes of complexity in intermittent time series segments.
Figures
Figures from the paper (3 more)
Reference graph
Works this paper leans on
-
[1]
Acharya U., R., O. Faust, N. Kannathal, T. Chua, and S. Laxminarayan (2005). Non-linear analysis of EEG signals at various sleep stages.Computer Methods and Programs in Biomedicine 80(1), 37–45. Acharya U, R., K. N, O. W. Sing, L. Y. Ping, and T. Chua (2004). Heart rate analysis in normal subjects of various age groups.BioMedical Engineering OnLine 3,
work page 2005
-
[3]
Again, 52 new time series are generated from the latest seasonal ARIMA processes
We obtain a seasonal ARIMA(p, d, q) process for each group. Again, 52 new time series are generated from the latest seasonal ARIMA processes. Then we regard them as observations and apply ourE to these new observations to check whether our approach can detect the change-points correctly. First, we need to estimate three seasonal ARIMA processes. For simpl...
work page 2013
-
[4]
ρp = 1 ρ1 ρ2 · · ·ρp−1 ρ1 1 ρ2 · · ·ρp−2 ρ2 ρ1 1 · · ·ρp−3
According to the Yule-Walker equations, ρ1 ρ2 ρ3 ... ρp = 1 ρ1 ρ2 · · ·ρp−1 ρ1 1 ρ2 · · ·ρp−2 ρ2 ρ1 1 · · ·ρp−3 ... ... ... ... ... ρp−1 ρp−2 ρp−3 · · · 1 ϕ1 ϕ2 ϕ3 ... ϕp , the autocorrelation ρi, i= 1 , 2, . . . ,∞ are totally determined by coefficient ϕk, k= 1 , . . . , p. Hence, Is(xm+1) is a functi...
work page 1995
-
[5]
This proof skill is frequently adopted in discussion of BIC or AIC consistency, see, for example, Shibata (1981, p.46), Vieu (1995,
work page 1981
-
[8]
(25) Let ˆm = arg min σ2 λ(m), we haveσ2 0( ˆmλ)/σ2 0(m0) → 1 almost surely
Based on Lemma C.1, previous discussion and Condition (C7), we have max m=1,...,M σ2 λ(m) − ˆσ2(m) σ2 0(m) = log(n) n 4 4+m ≤ log(n) n 4 4+O( √ log(n)) = o(1), a.s. (25) Let ˆm = arg min σ2 λ(m), we haveσ2 0( ˆmλ)/σ2 0(m0) → 1 almost surely. Because the previous results are almost surely convergence, so P ( ˆm = m0) → 1, holds as well, which completes the...
work page 1995
-
[9]
1 n nX i=1 P (i ̸∈ S(m)) # , (39) = Op
+ O(mn−1h) + Op(mn−1h− m+1 2 ), which immediately completes the proof comparing with ordern−1/2h−(m+1)/2. Lemma D.5. Given H0, under Assumptions 2.1 and 3.1, ifnhm/ log n → ∞, h → 0 and m < M. The order of the remainder term in equation(26) is Op n−3/2h−3m/2(log n)1/2 + m2h6 . Proof of Lemma D.5.Under Assumptions 2.1 and 3.1, we can obtain the uniform rat...
work page 2007
-
[10]
Forany x, y∈ [0, 1], denoteγ1(x, y) = R 1 0 [K J h (x∗, y)− R 1 0 K J h (x∗, y∗)g1(y∗)dy∗]dx∗/g1(x)
Proof of Lemma D.6.Firstly, we give a similar result for the univariate kernel; then we extend this result tothemultivariatekernel. Forany x, y∈ [0, 1], denoteγ1(x, y) = R 1 0 [K J h (x∗, y)− R 1 0 K J h (x∗, y∗)g1(y∗)dy∗]dx∗/g1(x). The numerator ofγ1(x, y) includes two terms. The first term can be expanded as Z 1 0 K J h (x∗ − y) dx∗ = Z h 0 K J h (x∗ − ...
work page 2005
-
[11]
− ˆC(m) − ˆcm = Op(n−1/2h4), this completes the proof. Proof of Theorem 3.1.Following the Theorems A.6 – A.9 in Hong and White (2005), one can extend their theory to multivariate U-statistics with boundedm. Hong and White (2005) discussed the pair variables Zjt = ( Xt, Xt−j)⊤ and j = o(n). Theorem A.6 constructs a new 2j-dependent process to show the depe...
work page 2005
Show all 13 references
-
[12]
The indices ofi, jhave at mostO(m) and O(m) choices respectively givenm < M
The number of pair(s, t), t − s ≥ m is of orderO(n2), for each given (s, t), ifZi;m has at least one overlap variable withZs;m or Zt;m, andZj;m has at least one overlap variable withZs;m or Zt;m as well, then the expectation is nonzero. The indices ofi, jhave at mostO(m) and O...
2005
-
[24]
gold standard
Braun, J. V., R. K. Braun, and H. G. Muller (2000). Multiple changepoint fitting via quasilikelihood, with application to DNA sequence segmentation.Biometrika 87(2), 301–314. Brown, B. M. (1971). Martingale central limit theorems.The Annals of Mathematical Statistics 42(1), 59...
2000
-
[314]
and Shao (1997, 232). Therefore, the sketch of our proof can be summarized into the following two steps: (1) We need to discuss the consistency of lag order selected viaˆσ2(m); (2) We extend the result to the penalty versionσ2 λ(m) with controlling λ(m, ˆhm) → 0 in an appropri...
1997
-
[568]
Shibata, R. (1981). An optimal selection of regression variables.Biometrika 68(1), 45–54. Vieu, P. (1991). Smoothing techniques in time series analysis. In G. Roussas (Ed.), Nonparamet- ric Functional Estimation and Related Topics, NATO ASI Series, pp. 271–283. Dordrecht: Spri...
1981
-
[1995]
Next, we introduce the conditions used in our proof
except the definitions ofˆσ2(m) and λ(m, ˆhm). Next, we introduce the conditions used in our proof. (C1) The time series {Xi}i∈N is α-mixing, the mixing coefficient α(n) satisfies: ∃ s > 0, ∃ 0 < t < 1, ∀ n ≥ 1, α(n) ≤ stn. (C2) For each 1 ≤ m < M, there exists the nonlinear a...
1995
Reviewed August 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.