A Frequentist Approach to Change Point Detection: Methods and Applications

Debanjana Datta

REVIEW 5 major objections 4 minor 24 references

This paper claims that a change point in a functional time series can be estimated by choosing the candidate split that maximizes the conditional probability of a change point given all other fitted parameters, and that this estimator is co

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

The paper proposes estimating the change point in functional time series by maximizing the conditional probability P(tau|others) under an FAR(1)/DLM model and claims consistency (proved only in a sketchy appendix).

T0 review reviewed 2026-08-02 challenge →

load-bearing objection The change-point estimator is not proven consistent: the appendix's key assumption Z_t'Z_t=I_M contradicts the sparse-data setting the paper advertises. the 5 major comments →

arxiv 2607.13852 v1 pith:S24TDJMS submitted 2026-07-15 math.ST stat.TH

A Frequentist Approach to Change Point Detection: Methods and Applications

Debanjana Datta This is my paper

classification math.ST stat.TH MSC 62R1062G2062M10

keywords functional datachange point detectionfunctional autoregressive processdynamic linear modelsparse and dense observation gridsmean shiftvolatility changeconsistency

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 proposes a frequentist procedure for detecting a single change point in a functional time series that may be observed on sparse or dense grids. The central claim is that the true change point can be recovered by estimating the model under each candidate split and choosing the split that maximizes the conditional probability of a change point given the fitted parameters. The author applies the same recipe to two problems: a shift in the mean function and a change in the measurement-error volatility. A consistency proof is sketched for the mean-shift case, and simulations are reported in which the estimated change point matches the true one in every tested configuration for the mean problem. If correct, the method offers a single algorithmic answer to a problem that has usually been handled by separate tests or Bayesian priors.

Core claim

The paper's central claim is that the change point estimator—defined as the τ that maximizes the conditional probability [τ|...] after estimating all model parameters—is consistent: as the sample size grows, the selected split concentrates on the true change point. The claim is made for both mean-shift and volatility-change versions of a functional autoregressive model expressed as a dynamic linear model. The proposed algorithm alternates between estimating the functional mean, the autoregressive operator, and the error variance under each candidate τ, then selects τ by the conditional-probability criterion.

What carries the argument

The load-bearing construction is the dynamic linear model representation y_t = Z_t μ + Z_t α_t + ν_t, α_t = Ψ Q α_{t−1} + ε_t, where Z_t is an incidence matrix encoding which grid points of a common evaluation set are observed at time t, Ψ is an autoregressive integral operator discretized with quadrature, and Q is a quadrature weight matrix. The estimator is the candidate τ maximizing the conditional probability [τ|...]; the consistency appendix derives squared-error expressions for the estimated mean functions under true and estimated change points. The proof's algebra relies on the orthonormality condition Z_t'Z_t = I_M for every t, which holds only when each time point's observations for

Load-bearing premise

The consistency proof assumes Z_t'Z_t = I_M for every time point, meaning each observed set of points acts as an orthonormal projection; for sparse functional data on irregular grids, Z_t'Z_t is a diagonal matrix of measurement counts, so the key algebra in the proof does not apply to the very setting the paper emphasizes.

What would settle it

Run the proposed algorithm on a simulated sparse functional time series where the number and locations of observations differ across curves, so that Z_t'Z_t is diagonal with unequal entries, and check whether the selected change point converges to the true τ as T grows; if it does not concentrate on τ, the consistency claim as stated fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

For dense functional data, the method gives a single estimator for mean shifts, and the simulation section reports exact recovery of the change point in every tested configuration from T=50 to T=200.
For volatility changes, increasing the number of measurements per curve improves the accuracy of the estimated change point, indicating the procedure exploits within-curve information.
If the consistency claim is accepted, the conditional-probability criterion becomes a general frequentist rule for single change-point detection in functional autoregressive time series, covering both sparse and dense observation designs.
The real-data examples suggest the estimator can be applied to long temperature records and to financial price series, where a change point corresponds to a historical climate regime or a market volatility shift.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The conditional-probability criterion is only defined procedurally; an explicit formula for [τ|...] would be needed to turn the algorithm into a fully reproducible estimator, and a direct derivation of its distribution is not given in the paper.
The proof's orthonormality assumption suggests the consistency result, as written, may hold only for balanced or dense designs; probing the method on unbalanced sparse grids would separate the paper's claim from its proof.
A natural extension the author leaves implicit is sequential or multiple change-point detection, where the same conditional-probability maximization could be applied recursively.
The method's performance likely depends on the choice of the common evaluation grid and the B-spline basis used for Ψ, since those choices determine the approximation error in the state-space model.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

The change-point estimator is not proven consistent: the appendix's key assumption Z_t'Z_t=I_M contradicts the sparse-data setting the paper advertises.

read the letter

Dear colleague,

Quick take: this paper has a plausible idea - use the conditional probability of the change point in a DLM for FAR(1) functional data, maximize it, and estimate both mean and volatility shifts. But the central consistency claim, stated in the abstract, is not supported. The appendix proof breaks down exactly where it needs to work.

What's new: the estimator itself is a reasonable construction. The paper correctly connects functional autoregression with a dynamic linear model (following Kowal et al.) and designs an algorithm that is meant to handle sparse data. The simulations, while thin, show the method can recover the true change point in some settings. The literature review is broad and relevant.

The problem is the proof. The appendix never defines the conditional probability [tau|...] that Section 4 says is maximized; the algebra is about least-squares mean estimates, not about the argmax of a posterior-like objective. More seriously, the key step sets Z_t^T Z_t = I_M for all t. In the model of Section 3, Z_t is an m_t x M incidence matrix. For any sparse design with m_t < M, Z_t^T Z_t is a diagonal matrix of grid-point counts, not the identity. So the displayed equality - and the resulting ||B theta_hat_1 - B theta_hat_1^E||^2 = ((tau_hat - tau)/tau_hat)^2 ||mu_2 - mu_1||^2 - is false in the very sparse regime the paper claims to handle. The proof also stops without showing that the argmax of [tau|...] converges. There is no proof at all for the volatility case.

The manuscript also reads as an early draft: missing definitions, garbled equations, incomplete references (the reference list literally has a line 'Berkes et al. Banerjee and Mazumder'), and no code. The simulations are reported without error bars.

For me this is a desk reject rather than a referee assignment: the central claim is invalid as written, and the flaw is load-bearing, not cosmetic. A reader interested in change point detection for functional data might pick up the algorithm as a heuristic, but would get no reliable theory from it. The idea could become a legitimate contribution if the estimator is fully specified, the proof is redone without the orthonormality assumption, and the volatility case is handled. As is, I wouldn't cite it, and I'd be surprised if a serious referee could do anything but ask for major revision. I would not send it to peer review in its current form.

Referee Report

5 major / 4 minor

Summary. The paper proposes a functional-autoregressive dynamic linear model for functional time series observed on sparse or dense grids and develops algorithms for detecting a change point in the mean function or in the measurement-error volatility. The estimator is defined in Section 4 as the change point τ that maximizes an unspecified conditional probability [τ|...], after least-squares/spline estimation of the other model components. The abstract states that the estimator has been proved consistent, and the Appendix attempts such a proof for the mean-shift problem. Simulation tables and two real-data applications are presented. The core contribution, as advertised, is the consistency result; that result is not established by the manuscript.

Significance. If the estimator were fully specified and its consistency rigorously proven, the paper would address a useful gap: change point detection for functional time series with sparse observation supports and with both mean and volatility shifts. The dynamic linear model representation and the practical algorithms are plausible starting points. However, the manuscript as written does not supply a well-defined estimator, and the only proof supporting the central claim breaks down exactly in the sparse regime that is part of the advertised contribution. The numerical examples are illustrative but not validated with uncertainty measures, replication, or a clear comparison protocol. The paper is therefore not ready for publication; its main claim is unsupported.

major comments (5)

[Section 4 and Appendix] The estimator is never precisely defined. Section 4.1 and 4.2 instruct the reader to 'Find that particular value of τ which maximizes the conditional probability [τ|...]', but no likelihood, prior, or full conditional distribution is written anywhere. The Appendix analyzes least-squares estimates θ̂1, θ̂2 under τ and τ̂ and derives discrepancies; it never shows that the argmax of [τ|...] is related to those least-squares quantities. Consequently, the abstract's claim that 'it has been proved that the proposed estimator is consistent' has no support: there is no theorem statement and no proof connecting the actual objective to the algebra in the Appendix.
[Appendix, 'Consider, Z_t^T Z_t = I_M ∀t'] The key simplification in the proof assumes Z_t^T Z_t = I_M for every t. In Section 3, Z_t is the m_t × M incidence matrix of observation points, so for genuinely sparse observation schemes with m_t < M, Z_t^T Z_t is a diagonal matrix whose diagonal entries are 0/1 counts and whose trace is m_t. It cannot equal the identity unless every curve is observed at all M grid points. Therefore the displayed equalities ||(A_1^{*T}A_1^*)^{-1}A_1^{*T}Z(θ2−θ1)||_2^2 = ((τ̂−τ)/τ̂)^2||μ2−μ1||_2^2 and the analogous one for A_2 are false in the sparse regime advertised in the abstract and simulations. The central proof collapses exactly where the method is supposed to apply.
[Appendix vs Section 4.1] The proof is not for the algorithm that is stated. Section 4.1 says 'Estimate μ1 and μ2 by the method of least squares', whereas the Appendix assumes the parametric form μ1 = Bθ1, μ2 = Bθ2 and performs spline-constrained least squares. If the algorithm is to be interpreted with B-spline mean functions, that must be stated explicitly; otherwise the proof concerns a different estimator. In addition, the limits θ̂1 → θ1 and θ̂E1 → θ1 −(A_1^{*T}A_1^*)^{-1}A_1^{*T}Z(θ2−θ1) are asserted without conditions on the random τ̂, the growth of the number of basis functions, or the observation design; these are exactly the facts that need to be proved.
[Section 4.2 and Appendix] There is no consistency argument for the volatility change point. The Appendix's entire computation concerns a mean shift. For the volatility problem, Section 4.2 estimates σ1^2 and σ2^2 for each τ and again selects τ by the unspecified [τ|...], but no residual-sum-of-squares expansion under a wrong τ for a variance change is given. The claim that the proposed estimator is consistent therefore has no support for the second problem treated in the paper.
[Sections 3 and 4 and Appendix] The proof silently assumes that the latent autoregressive process α_t is identifiable from y_t and recoverable by the spline step. Since y_t has only m_t coordinates while α_t is M-dimensional, for m_t < M the likelihood is flat in the unobserved coordinates unless strong identifiability constraints are imposed. The Appendix's expansions σ̂ν^E = σ̂ν^2 + K + A_1 or +A_2 presuppose that the spline/DLM estimates ˆα_t and Ψ converge to their true counterparts. No conditions on the number and placement of observation points, the B-spline basis, or the quadrature approximation are stated, so the proof does not cover the sparse functional data setting the paper claims to address.

minor comments (4)

[Section 5, Tables 1 and 2] Simulation reporting is incomplete. There are no replication counts, standard errors, or mean absolute errors, and some entries are difficult to interpret; e.g., in Table 1 with T=50, m=30, the 'CP at T/2' entry is 17 while the text says the method gives accurate results. The statement in Section 5 that 'the estimated change points exactly matches with the true ones for all sample sizes' for the mean problem is not backed by any table.
[Section 6, Table 3] Table 3 is labeled 'Performance comparison' but has no column headers identifying the methods, no uncertainty measures, and no discussion of the discrepancy among the three estimated change points. The comparison with Banerjee and Mazumder [2018] and Berkes et al. [2009] is therefore not quantitative.
[Notation and algorithm] The notation [τ|...] is informal. The paper should write the conditioning variables explicitly, e.g., p(τ | y, μ, α, σ^2, Ψ), and state the prior on τ. Also, the iterative loop in Section 4.1 that re-estimates α_t, μ, and σ^2 gives no convergence criterion; the algorithm is not fully specified.
[References] The reference list contains broken entries: 'Berkes et al.' and 'Banerjee and Mazumder' appear as placeholder fragments; Aue and Van Delft [2020] lacks venue details. These need to be completed.

Circularity Check

0 steps flagged

No significant circularity: the consistency proof is invalid for sparse designs, but its flaw is an invalid assumption, not a reduction of the conclusion to its inputs.

full rationale

The paper's central claim is that the proposed estimator is consistent. The Appendix attempts a proof by comparing least-squares mean estimates under the true and estimated change points, and it invokes 'Consider, Z_t^T Z_t = I_M ∀t' to collapse certain block-matrix expressions. In the Section 3 model, Z_t is an m_t × M incidence matrix, so for a genuinely sparse design Z_t^T Z_t is a diagonal matrix of observation counts, not the identity; the displayed algebra therefore does not hold in the setting the paper advertises. However, this is an invalid or incomplete proof, not circularity: no equation in the paper is shown to be equivalent to its own input by construction, and no fitted parameter is renamed as a prediction. The estimator is defined as the argmax of an unspecified conditional probability [τ|...], and the Appendix never demonstrates that this argmax converges; that is a missing justification, not a self-referential reduction. The volatility change-point claim has no proof at all, which is a correctness and completeness concern, not a circularity concern. The paper's literature-review citations are external prior work rather than a load-bearing self-citation chain. Accordingly, no load-bearing circular step can be exhibited, and the circularity score is 0 despite serious mathematical gaps.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The ledger shows the paper's load-bearing commitments: a Gaussian FAR(1) DLM, an unspecified conditional probability to maximize, and a proof condition (Z_t'Z_t = I_M) that contradicts the sparse-data motivation. No new physical entities are invented.

free parameters (2)

C_psi (kernel squared norm) = 0.4
Set by hand in Section 5 to ensure stationarity; choice not justified by data.
sigma_nu measurement-error std = 0.001
Fixed simulation input, not a fitted parameter.

axioms (4)

domain assumption FAR(1) integral operator model with Gaussian noise
The entire model, Section 3, restricts to integral operators and Gaussian noise.
ad hoc to paper The estimator is defined as maximizer of [tau|...] without specifying a derived conditional distribution
Section 4 says 'Find that particular value of tau which maximizes the conditional probability [tau|...]', but no explicit formula for this probability appears; the proof never derives it.
ad hoc to paper Z_t' Z_t = I_M in the consistency proof
Appendix, 'Consider, Z_t^T Z_t = I_M forall t', which is false for realistic sparse designs and for the manuscript's own m_t x M incidence matrices with m_t < M.
ad hoc to paper Spline fits recover alpha_t and Psi separately
Algorithm 4.1 fits a spline to {y_t - mu_hat} to obtain {alpha_hat_t}, then estimates Psi from these; no identifiability or convergence guarantee is given.

reviewed 2026-08-02 · how reviews work

0 comments

Cite this review

Pith. "Pith review of A Frequentist Approach to Change Point Detection: Methods and Applications." pith.science (2026). https://pith.science/paper/S24TDJMS

@misc{pith2026260713852,
  author       = {Pith},
  title        = {Pith review of: A Frequentist Approach to Change Point Detection: Methods and Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S24TDJMS}},
  note         = {Machine review of arXiv:2607.13852}
}

Share X Bluesky LinkedIn Reddit HN

read the original abstract

In this paper we study the problem of change point detection in functional time series where the observations are allowed to vary on both sparse and dense support. We address the problem of mean shift as well as the process volatility. Our methodology is based on the maximization of the conditional probability of change point given all the other parameters. Further, it has been proved that the proposed estimator is consistent.

discussion (0)

Reference graph

Works this paper leans on

24 extracted references

[1]

Journal of Business & Economic Statistics , volume=

Functional autoregression for sparsely sampled data , author=. Journal of Business & Economic Statistics , volume=. 2019 , publisher=

2019
[2]

Journal of Multivariate Analysis , volume=

Testing the stability of the functional autoregressive process , author=. Journal of Multivariate Analysis , volume=. 2010 , publisher=

2010
[3]

Journal of the American Statistical Association , volume=

Portmanteau test of independence for functional observations , author=. Journal of the American Statistical Association , volume=. 2007 , publisher=

2007
[4]

Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

Thin plate regression splines , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2003 , publisher=

2003
[5]

Journal of Quantitative Economics , volume=

Bayesian testing of granger causality in functional time series , author=. Journal of Quantitative Economics , volume=. 2022 , publisher=

2022
[6]

2017 , publisher=

Introduction to functional data analysis , author=. 2017 , publisher=

2017
[7]

2000 , publisher=

Time series analysis and its applications , author=. 2000 , publisher=

2000
[8]

2012 , publisher=

Time series analysis by state space methods , author=. 2012 , publisher=

2012
[9]

Journal of Econometrics , volume=

Testing stationarity of functional time series , author=. Journal of Econometrics , volume=. 2014 , publisher=

2014
[10]

Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

Detecting changes in the mean of functional observations , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2009 , publisher=

2009
[11]

Journal of Multivariate Analysis , volume=

Estimation of a change-point in the mean function of functional data , author=. Journal of Multivariate Analysis , volume=. 2009 , publisher=

2009
[12]

The Annals of Statistics , volume=

Break detection in the covariance structure of multivariate time series models , author=. The Annals of Statistics , volume=
[13]

New York , year=

Functional Data Analysis 2 edition Springer , author=. New York , year=
[14]

2000 , publisher=

Linear processes in function spaces: theory and applications , author=. 2000 , publisher=

2000
[15]

Annals of the Institute of Statistical Mathematics , volume=

A more powerful test identifying the change in mean of functional data , author=. Annals of the Institute of Statistical Mathematics , volume=. 2018 , publisher=

2018
[16]

Testing for stationarity of functional time series in the frequency domain , author=
[17]

Statistica Sinica , volume=

A nonparametric test for stationarity in functional time series , author=. Statistica Sinica , volume=. 2021 , publisher=

2021
[18]

Journal of the royal statistical society: series C (applied statistics) , volume=

Hierarchical Bayesian analysis of changepoint problems , author=. Journal of the royal statistical society: series C (applied statistics) , volume=. 1992 , publisher=

1992
[19]

The Annals of Mathematical Statistics , volume=

Estimating the current mean of a normal distribution which is subjected to changes in time , author=. The Annals of Mathematical Statistics , volume=. 1964 , publisher=

1964
[20]

Biometrika , volume=

A Bayesian approach to inference about a change-point in a sequence of random variables , author=. Biometrika , volume=. 1975 , publisher=

1975
[21]

Journal of Statistical Planning and Inference , volume=

Bayesian change point detection for functional data , author=. Journal of Statistical Planning and Inference , volume=. 2021 , publisher=

2021
[22]

Journal of Statistical Software , volume=

Bayesian analysis for penalized spline regression using WinBUGS , author=. Journal of Statistical Software , volume=
[23]

2006 , publisher=

Bayesian forecasting and dynamic models , author=. 2006 , publisher=

2006
[24]

Communications in Statistics-Theory and Methods , volume=

Bayesian approach to change point problems , author=. Communications in Statistics-Theory and Methods , volume=. 1997 , publisher=

1997

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.

[1] [1]

Journal of Business & Economic Statistics , volume=

Functional autoregression for sparsely sampled data , author=. Journal of Business & Economic Statistics , volume=. 2019 , publisher=

2019

[2] [2]

Journal of Multivariate Analysis , volume=

Testing the stability of the functional autoregressive process , author=. Journal of Multivariate Analysis , volume=. 2010 , publisher=

2010

[3] [3]

Journal of the American Statistical Association , volume=

Portmanteau test of independence for functional observations , author=. Journal of the American Statistical Association , volume=. 2007 , publisher=

2007

[4] [4]

Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

Thin plate regression splines , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2003 , publisher=

2003

[5] [5]

Journal of Quantitative Economics , volume=

Bayesian testing of granger causality in functional time series , author=. Journal of Quantitative Economics , volume=. 2022 , publisher=

2022

[6] [6]

2017 , publisher=

Introduction to functional data analysis , author=. 2017 , publisher=

2017

[7] [7]

2000 , publisher=

Time series analysis and its applications , author=. 2000 , publisher=

2000

[8] [8]

2012 , publisher=

Time series analysis by state space methods , author=. 2012 , publisher=

2012

[9] [9]

Journal of Econometrics , volume=

Testing stationarity of functional time series , author=. Journal of Econometrics , volume=. 2014 , publisher=

2014

[10] [10]

Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

Detecting changes in the mean of functional observations , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2009 , publisher=

2009

[11] [11]

Journal of Multivariate Analysis , volume=

Estimation of a change-point in the mean function of functional data , author=. Journal of Multivariate Analysis , volume=. 2009 , publisher=

2009

[12] [12]

The Annals of Statistics , volume=

Break detection in the covariance structure of multivariate time series models , author=. The Annals of Statistics , volume=

[13] [13]

New York , year=

Functional Data Analysis 2 edition Springer , author=. New York , year=

[14] [14]

2000 , publisher=

Linear processes in function spaces: theory and applications , author=. 2000 , publisher=

2000

[15] [15]

Annals of the Institute of Statistical Mathematics , volume=

A more powerful test identifying the change in mean of functional data , author=. Annals of the Institute of Statistical Mathematics , volume=. 2018 , publisher=

2018

[16] [16]

Testing for stationarity of functional time series in the frequency domain , author=

[17] [17]

Statistica Sinica , volume=

A nonparametric test for stationarity in functional time series , author=. Statistica Sinica , volume=. 2021 , publisher=

2021

[18] [18]

Journal of the royal statistical society: series C (applied statistics) , volume=

Hierarchical Bayesian analysis of changepoint problems , author=. Journal of the royal statistical society: series C (applied statistics) , volume=. 1992 , publisher=

1992

[19] [19]

The Annals of Mathematical Statistics , volume=

Estimating the current mean of a normal distribution which is subjected to changes in time , author=. The Annals of Mathematical Statistics , volume=. 1964 , publisher=

1964

[20] [20]

Biometrika , volume=

A Bayesian approach to inference about a change-point in a sequence of random variables , author=. Biometrika , volume=. 1975 , publisher=

1975

[21] [21]

Journal of Statistical Planning and Inference , volume=

Bayesian change point detection for functional data , author=. Journal of Statistical Planning and Inference , volume=. 2021 , publisher=

2021

[22] [22]

Journal of Statistical Software , volume=

Bayesian analysis for penalized spline regression using WinBUGS , author=. Journal of Statistical Software , volume=

[23] [23]

2006 , publisher=

Bayesian forecasting and dynamic models , author=. 2006 , publisher=

2006

[24] [24]

Communications in Statistics-Theory and Methods , volume=

Bayesian approach to change point problems , author=. Communications in Statistics-Theory and Methods , volume=. 1997 , publisher=

1997

A Frequentist Approach to Change Point Detection: Methods and Applications

The reading

Core claim §

What carries the argument

Load-bearing premise §

What would settle it §

If this is right

Where Pith is reading between the lines

discussion (0)

Core claim

Load-bearing premise

What would settle it