pith. sign in

arxiv: 2509.07112 · v3 · submitted 2025-09-08 · 🧮 math.ST · stat.ME· stat.TH

Self-Normalization for CUSUM-based Change Detection in Locally Stationary Time Series

Florian Heinrichs This is my paper

Pith reviewed 2026-05-18 17:32 UTC · model grok-4.3

classification 🧮 math.ST stat.MEstat.TH
keywords self-normalizationCUSUM testchange detectionlocally stationary time seriespartial sum processBrownian sheetmean change
0
0 comments X

The pith

A bivariate partial sum process enables self-normalized CUSUM tests for mean changes in locally stationary time series without estimating time-varying variance.

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

The paper introduces a bivariate partial sum process for locally stationary time series and proves its weak convergence to a Brownian sheet. This construction supports self-normalized CUSUM statistics that reach the nominal asymptotic level under no change and remain consistent against abrupt, gradual, and multiple mean shifts. Standard self-normalization works for stationary series because a constant long-run variance factors out, but local stationarity produces an integrated stochastic variance that blocks this cancellation. The bivariate approach overcomes the obstacle by pairing processes so the variance terms cancel in the limit. Simulations indicate accurate size and better finite-sample power than methods that estimate the long-run variance explicitly.

Core claim

A bivariate partial sum process is constructed for locally stationary time series and shown to converge weakly to a Brownian sheet. This limit permits self-normalized CUSUM test statistics for the mean that attain asymptotic level alpha under the null of no change and are consistent against abrupt, gradual, and multiple changes under mild assumptions on the local stationarity and the change structure.

What carries the argument

The bivariate partial sum process, which augments the usual partial sums with an auxiliary process so that the time-varying long-run variance cancels in the limiting distribution.

If this is right

  • The tests maintain correct asymptotic size under the null of no change for locally stationary series.
  • The statistics are consistent against single abrupt mean changes.
  • The statistics are consistent against gradual mean changes.
  • The statistics are consistent against multiple mean changes.
  • Finite-sample simulations show accurate size and higher power than variance-estimation competitors.

Where Pith is reading between the lines

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

  • The same bivariate construction might be adapted to detect changes in other functionals such as variance or autocovariance if suitable auxiliary processes are identified.
  • Extension to multivariate locally stationary series would require a corresponding matrix-valued Brownian-sheet limit.
  • The method offers a practical route for change detection in economic or environmental series whose second-order structure evolves slowly over time.
  • Direct comparison of computational cost against plug-in long-run-variance estimators on large data sets would clarify when self-normalization is preferable.

Load-bearing premise

Local stationarity produces a partial-sum limit that is an integral of a time-varying volatility against Brownian motion, so the usual constant-variance factorization cannot be used for normalization.

What would settle it

A simulation or analytic counterexample in which the self-normalized statistic fails to converge in distribution to the claimed Brownian-sheet functional under a specific locally stationary null process with non-constant sigma(x) would falsify the asymptotic level claim.

Figures

Figures reproduced from arXiv: 2509.07112 by Florian Heinrichs.

Figure 1
Figure 1. Figure 1: Visualization of indices of the bivariate partial sum process [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Various mean functions, used to generate time series under the alternative. [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Exemplary trajectories of AR(1) processes with variance σ2 (left) and σ3 (right), for cσ = 1, µ(x) ≡ 0 and n = 200. Exemplary trajectories of an AR(1) process for σ2 and σ3 under H0, for the constant mean function µ(x) ≡ 0, are displayed in [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Mean temperatures in Gayndah (left), Robe (center) and Sydney (right) for the [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
read the original abstract

A new bivariate partial sum process for locally stationary time series is introduced and its weak convergence to a Brownian sheet is established. This construction enables the development of a novel self-normalized CUSUM test statistic for detecting changes in the mean of a locally stationary time series. For stationary data, self-normalization relies on the factorization of a constant long-run variance and a stochastic factor. In this case, the CUSUM statistic can be divided by another statistic proportional to the long-run variance, so that the latter cancels, avoiding estimation of the long-run variance. Under local stationarity, the partial sum process converges to $\int_0^t \sigma(x) d B_x$ and no such factorization is possible. To overcome this obstacle, a bivariate partial-sum process is introduced, allowing the construction of self-normalized test statistics under local stationarity. Weak convergence of the process is proven, and it is shown that the resulting self-normalized tests attain asymptotic level $\alpha$ under the null hypothesis of no change, while being consistent against abrupt, gradual, and multiple changes under mild assumptions. Simulation studies show that the proposed tests have accurate size and substantially improved finite-sample power relative to existing approaches. Two data examples illustrate practical performance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 3 minor

Summary. The paper introduces a bivariate partial sum process for locally stationary time series and establishes its weak convergence to a Brownian sheet. This enables construction of self-normalized CUSUM test statistics for detecting changes in the mean that attain asymptotic level α under the null of no change and are consistent against abrupt, gradual, and multiple changes, without explicit estimation of the time-varying long-run variance. Simulations indicate accurate size and improved power relative to existing methods, with two real-data illustrations.

Significance. If the weak-convergence result holds, the work is significant because it resolves the factorization obstacle that prevents standard self-normalization under local stationarity, where the partial-sum limit is ∫σ(x)dB_x. By delivering a pivotal limiting distribution via the bivariate construction and continuous mapping, the approach avoids long-run-variance estimation that is especially difficult in this setting and yields consistency under a range of alternatives. The reported finite-sample gains and practical examples suggest immediate applicability in fields such as econometrics and signal processing.

major comments (2)
  1. [§3, Theorem 3.1] §3, Theorem 3.1: the weak convergence of the bivariate partial-sum process to the Brownian sheet is the load-bearing step for the pivotal limit; the proof sketch should explicitly verify that the covariance kernel of the bivariate object cancels the time-varying σ(x) factor uniformly in the two time arguments, rather than only pointwise.
  2. [§4.2] §4.2, the consistency argument under gradual changes: the rate at which the change magnitude must grow relative to the local-stationarity bandwidth is not stated quantitatively; without this, it is unclear whether the claimed consistency holds at the same rate as the abrupt-change case.
minor comments (3)
  1. Notation for the local-stationarity kernel and the two time indices of the bivariate process should be unified across the abstract, Section 2, and the theorems to avoid reader confusion.
  2. Figure 2 (power curves) lacks error bars or Monte-Carlo standard errors; adding them would strengthen the claim of “substantially improved” finite-sample power.
  3. The mixing and moment conditions (Assumptions A1–A3) are listed but their relation to existing local-stationarity frameworks (e.g., Dahlhaus) is not compared; a short remark would help readers assess novelty.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive evaluation and the constructive comments, which help clarify key technical points in the proofs and consistency results. We address each major comment below.

read point-by-point responses

  1. Referee: §3, Theorem 3.1: the weak convergence of the bivariate partial-sum process to the Brownian sheet is the load-bearing step for the pivotal limit; the proof sketch should explicitly verify that the covariance kernel of the bivariate object cancels the time-varying σ(x) factor uniformly in the two time arguments, rather than only pointwise.

    Authors: We agree that making the uniform cancellation explicit strengthens the presentation. The bivariate partial-sum process is constructed so that its covariance kernel takes the product form min(s1,s2)min(t1,t2) after integration against the slowly varying σ(x). In the existing proof of Theorem 3.1, finite-dimensional convergence is obtained by direct computation of the covariance, and the local-stationarity condition ensures the error from replacing σ(x) by a constant over small intervals vanishes. To address the referee’s request, we will revise the proof sketch to include an explicit uniform bound: for any ε>0 there exists δ>0 such that the supremum over all pairs (s1,t1),(s2,t2) of the difference between the actual covariance and the Brownian-sheet kernel is controlled by the modulus of continuity of σ and the bandwidth, uniformly on the grid. This confirms the cancellation holds uniformly in both time arguments. revision: yes

  2. Referee: §4.2, the consistency argument under gradual changes: the rate at which the change magnitude must grow relative to the local-stationarity bandwidth is not stated quantitatively; without this, it is unclear whether the claimed consistency holds at the same rate as the abrupt-change case.

    Authors: We thank the referee for noting this omission. The consistency proof for gradual changes in §4.2 proceeds by approximating the smooth transition by locally stationary blocks of width proportional to the bandwidth h_n and showing that the self-normalized CUSUM diverges whenever the cumulative mean shift exceeds the stochastic fluctuation order. While the paper states consistency under mild assumptions that implicitly cover both abrupt and gradual cases, we acknowledge that an explicit rate comparison is useful. In the revision we will add a remark stating that, under the maintained bandwidth conditions (h_n→0, nh_n→∞), consistency holds for gradual changes as soon as the change magnitude δ_n satisfies δ_n/h_n→∞, which is the natural analogue of the abrupt-change requirement δ√n→∞ and therefore preserves the same asymptotic rate up to the bandwidth factor. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper defines a new bivariate partial-sum process for locally stationary series, proves its weak convergence to a Brownian sheet via standard arguments under local stationarity and weak dependence, and applies the continuous-mapping theorem to obtain a pivotal limiting distribution for the self-normalized CUSUM. This construction cancels the time-varying long-run variance factor without estimation or fitting. No step reduces a claimed prediction or test statistic to a fitted input by construction, no load-bearing self-citation chain appears, and the central results follow directly from the introduced object and established weak-convergence tools rather than from re-labeling or self-referential definitions.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The contribution rests on the local-stationarity framework and the new bivariate process; no free parameters are fitted inside the central claim, and the main background assumptions are standard domain conditions for weak convergence.

axioms (1)
  • domain assumption The time series is locally stationary with the regularity conditions that make the partial-sum process converge to ∫_0^t σ(x) dB_x
    Invoked to explain why ordinary self-normalization fails and to justify the bivariate construction.
invented entities (1)
  • bivariate partial sum process no independent evidence
    purpose: To enable cancellation of the time-varying stochastic factor so that self-normalized test statistics can be formed
    Newly defined in the paper to overcome the absence of factorization under local stationarity.

pith-pipeline@v0.9.0 · 5748 in / 1450 out tokens · 50891 ms · 2026-05-18T17:32:38.905765+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages

  1. [1]

    Afonso, T. V. and Heinrichs, F. (2025). Consumer-grade eeg-based eye tracking. arXiv preprint arXiv:2503.14322

    work page arXiv 2025

  2. [2]

    Aston, J. A. D. and Kirch, C. (2012). Evaluating stationarity via change-point alternatives with applications to fMRI data . The Annals of Applied Statistics , 6(4):1906 -- 1948

    work page 2012

  3. [3]

    and Horv \'a th, L

    Aue, A. and Horv \'a th, L. (2013). Structural breaks in time series. Journal of Time Series Analysis , 34(1):1--16

    work page 2013

  4. [4]

    Baranowski, R., Chen, Y., and Fryzlewicz, P. (2019). Narrowest-over-threshold detection of multiple change points and change-point-like features. Journal of the Royal Statistical Society Series B: Statistical Methodology , 81(3):649--672

    work page 2019

  5. [5]

    Birr, S., Volgushev, S., Kley, T., Dette, H., and Hallin, M. (2017). Quantile spectral analysis for locally stationary time series. Journal of the Royal Statistical Society Series B: Statistical Methodology , 79(5):1619--1643

    work page 2017

  6. [6]

    B \"u cher, A., Dette, H., and Heinrichs, F. (2020). Detecting deviations from second-order stationarity in locally stationary functional time series. Annals of the Institute of Statistical Mathematics , 72(4):1055--1094

    work page 2020

  7. [7]

    B \"u cher, A., Dette, H., and Heinrichs, F. (2021). Are deviations in a gradually varying mean relevant? a testing approach based on sup-norm estimators. The Annals of Statistics , 49(6):3583--3617

    work page 2021

  8. [8]

    B \"u cher, A., Dette, H., and Heinrichs, F. (2023). A portmanteau-type test for detecting serial correlation in locally stationary functional time series. Statistical Inference for Stochastic Processes , 26(2):255--278

    work page 2023

  9. [9]

    and Jennessen, T

    B \"u cher, A. and Jennessen, T. (2024). Statistics for heteroscedastic time series extremes. Bernoulli , 30(1):46--71

    work page 2024

  10. [10]

    and Graham, M

    Chakraborti, S. and Graham, M. A. (2019). Nonparametric (distribution-free) control charts: An updated overview and some results. Quality Engineering , 31(4):523--544

    work page 2019

  11. [11]

    and Fryzlewicz, P

    Cho, H. and Fryzlewicz, P. (2015). Multiple-change-point detection for high dimensional time series via sparsified binary segmentation. Journal of the Royal Statistical Society Series B: Statistical Methodology , 77(2):475--507

    work page 2015

  12. [12]

    and Kirch, C

    Cho, H. and Kirch, C. (2024). Data segmentation algorithms: Univariate mean change and beyond. Econometrics and Statistics , 30:76--95

    work page 2024

  13. [13]

    Collins, D., Della-Marta, P., Plummer, N., and Trewin, B. (2000). Trends in annual frequencies of extreme temperature events in australia. Australian Meteorological Magazine , 49(4):277--292

    work page 2000

  14. [14]

    Dahlhaus, R. (1996). On the kullback-leibler information divergence of locally stationary processes. Stochastic processes and their applications , 62(1):139--168

    work page 1996

  15. [15]

    and Wu, W

    Dette, H. and Wu, W. (2019). Detecting relevant changes in the mean of nonstationary processes—a mass excess approach. The Annals of Statistics , 47(6):3578--3608

    work page 2019

  16. [16]

    Frick, K., Munk, A., and Sieling, H. (2014). Multiscale change point inference. Journal of the Royal Statistical Society Series B: Statistical Methodology , 76(3):495--580

    work page 2014

  17. [17]

    Fryzlewicz, P. (2018). Tail-greedy bottom-up data decompositions and fast multiple change-point detection . The Annals of Statistics , 46(6B):3390 -- 3421

    work page 2018

  18. [18]

    Gao, Z., Shang, Z., Du, P., and Robertson, J. L. (2019). Variance change point detection under a smoothly-changing mean trend with application to liver procurement. Journal of the American Statistical Association

    work page 2019

  19. [19]

    Heinrichs, F. (2023). Monitoring machine learning models: Online detection of relevant deviations. arXiv preprint arXiv:2309.15187

    work page arXiv 2023

  20. [20]

    and Dette, H

    Heinrichs, F. and Dette, H. (2021). A distribution free test for changes in the trend function of locally stationary processes. Electronic Journal of Statistics , 15(2):3762--3797

    work page 2021

  21. [21]

    Horv \'a th, L., Horv \'a th, Z., and Hu s kov \'a , M. (2008). Ratio tests for change point detection. In Beyond parametrics in interdisciplinary research: Festschrift in honor of Professor Pranab K. Sen , volume 1, pages 293--305. Institute of Mathematical Statistics

    work page 2008

  22. [22]

    Horv \'a th, L., Kokoszka, P., and Steinebach, J. (1999). Testing for changes in multivariate dependent observations with an application to temperature changes. Journal of Multivariate Analysis , 68(1):96--119

    work page 1999

  23. [23]

    M., Sieling, H., Polupanow, T., Diederichsen, U., Steinem, C., and Munk, A

    Hotz, T., Sch \"u tte, O. M., Sieling, H., Polupanow, T., Diederichsen, U., Steinem, C., and Munk, A. (2013). Idealizing ion channel recordings by a jump segmentation multiresolution filter. IEEE transactions on NanoBioscience , 12(4):376--386

    work page 2013

  24. [24]

    Jandhyala, V., Fotopoulos, S., MacNeill, I., and Liu, P. (2013). Inference for single and multiple change-points in time series. Journal of Time Series Analysis , 34(4):423--446

    work page 2013

  25. [25]

    R., Knight, R

    Karl, T. R., Knight, R. W., and Plummer, N. (1995). Trends in high-frequency climate variability in the twentieth century. Nature , 377(6546):217--220

    work page 1995

  26. [26]

    Khoshnevisan, D. (2006). Multiparameter processes: an introduction to random fields . Springer Science & Business Media

    work page 2006

  27. [27]

    Kirch, C., Muhsal, B., and Ombao, H. (2015). Detection of changes in multivariate time series with application to eeg data. Journal of the American Statistical Association , 110(511):1197--1216

    work page 2015

  28. [28]

    Kley, T., Volgushev, S., Dette, H., and Hallin, M. (2016). Quantile spectral processes: Asymptotic analysis and inference . Bernoulli , 22(3):1770 -- 1807

    work page 2016

  29. [29]

    Page, E. S. (1954). Continuous inspection schemes. Biometrika , 41(1/2):100--115

    work page 1954

  30. [30]

    and Rao, T

    Priestley, M. and Rao, T. S. (1969). A test for non-stationarity of time-series. Journal of the Royal Statistical Society Series B: Statistical Methodology , 31(1):140--149

    work page 1969

  31. [31]

    and Shao, X

    Rho, Y. and Shao, X. (2015). Inference for time series regression models with weakly dependent and heteroscedastic errors. Journal of Business & Economic Statistics , 33(3):444--457

    work page 2015

  32. [32]

    Shao, X. (2010). A self-normalized approach to confidence interval construction in time series. Journal of the Royal Statistical Society Series B: Statistical Methodology , 72(3):343--366

    work page 2010

  33. [33]

    Shao, X. (2015). Self-normalization for time series: a review of recent developments. Journal of the American Statistical Association , 110(512):1797--1817

    work page 2015

  34. [34]

    A., and Obimbo, C

    Sharma, S., Swayne, D. A., and Obimbo, C. (2016). Trend analysis and change point techniques: a survey. Energy, ecology and environment , 1(3):123--130

    work page 2016

  35. [35]

    Truong, C., Oudre, L., and Vayatis, N. (2020). Selective review of offline change point detection methods. Signal Processing , 167:107299

    work page 2020

  36. [36]

    Van Der Vaart, A. W. and Wellner, J. A. (1996). Weak Convergence and Empirical Processes . Springer

    work page 1996

  37. [37]

    Vogt, M. (2012). Nonparametric regression for locally stationary time series . The Annals of Statistics , 40(5):2601 -- 2633

    work page 2012

  38. [38]

    and Dette, H

    Vogt, M. and Dette, H. (2015). Detecting gradual changes in locally stationary processes . The Annals of Statistics , 43(2):713 -- 740

    work page 2015

  39. [39]

    Wolfe, D. A. and Schechtman, E. (1984). Nonparametric statistical procedures for the changepoint problem. Journal of Statistical Planning and Inference , 9(3):389--396

    work page 1984

  40. [40]

    Woodall, W. H. and Montgomery, D. C. (2014). Some current directions in the theory and application of statistical process monitoring. Journal of quality technology , 46(1):78--94

    work page 2014

  41. [41]

    and Zhou, Z

    Wu, W. and Zhou, Z. (2024). Multiscale jump testing and estimation under complex temporal dynamics. Bernoulli , 30(3):2372--2398

    work page 2024

  42. [42]

    and Li, X

    Zhao, Z. and Li, X. (2013). Inference for modulated stationary processes. Bernoulli: official journal of the Bernoulli Society for Mathematical Statistics and Probability , 19(1):205

    work page 2013

  43. [43]

    and Wu, W

    Zhou, Z. and Wu, W. B. (2009). Local linear quantile estimation for nonstationary time series. The Annals of Statistics , pages 2696--2729

    work page 2009