acceptodds
Withdrawn from ICLR 2027

Detection and Localization of Multi-Type Parameter Changes in Non-Stationary Data

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

This paper studies a common but complex change-point setting in time series, where a smoothly evolving mean trend may coexist with abrupt shifts in both the mean and variance. To address this joint problem in such complex non-stationary data, we propose a novel iterative estimation procedure that alternates between smoothing-spline estimation of the mean function and likelihood-ratio tests for mean and variance change-points. The resulting method jointly estimates the mean change-point and the jump size of the mean shift, the two smooth mean functions before and after the change-point, the variance change-point, and the variances before and after the change. We derive the asymptotic null distributions of the proposed test statistics and establish convergence rates for the estimators of the smooth mean functions, change-point locations, mean shift size, and variance parameters. Simulation studies and a real-data application demonstrate that the proposed procedure is stable across a range of settings and provides accurate change-point detection together with reliable trend estimation. The proposed method is computationally tractable and has potential applications in finance, meteorology, sensor monitoring, economics, and industrial process control.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Parameter Changes detection, Non-Stationary Data, Hypothesis test in nonparametric smoothing, Smoothly evolving mean

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_k79uOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies change-point detection in univariate observations with a smoothly evolving mean, an abrupt mean change, and an abrupt variance change, where the two change-points may occur at different locations. It proposes an iterative procedure alternating between weighted smoothing-spline estimation and likelihood-based change-point tests. The procedure estimates the mean functions, change-point locations, mean jump size, and variance levels. The paper presents consistency results, convergence rates, and asymptotic null distributions. Validation includes simulations covering different change configurations and noise distributions, together with an application to battery manufacturing measurements.

Critical strengths
  1. A practically meaningful joint estimation problem. The paper accommodates smooth trends and separate mean and variance change-points. Distinguishing these effects is consequential because trend misspecification can confound change detection, while identifying the parameter that changes can improve the interpretation of manufacturing anomalies.

  2. An interpretable procedure with informative empirical diagnostics. The method jointly reports trends, jump sizes, locations, and variance levels. The simulations cover several relevant change configurations, and the appendix reports both detection power and null rejection rates. These diagnostics provide useful evidence about the procedure’s performance and limitations.

Critical weaknesses
  1. The central theoretical guarantees are not adequately established by the current proofs. In Appendix C.3, page 24, the proof obtains Sk−Sτ0=OP(r1n2)S_k-S_{\tau_0}=O_P(r_{1n}^2) and then concludes the claimed localization bound. An order bound does not establish that incorrect candidate locations have a larger objective with high probability; a uniform positive separation argument is missing. Additionally, the proof of Theorem 4.2 treats the transformed observations as independent Gaussian observations with unit variance, without adequately accounting for fitted-residual dependence, estimated variance parameters, or the candidate-dependent construction xi(k0)x_i^{(k_0)}. The transformation in Appendix C.5 also adds the estimated jump after τ1\tau_1, whereas Equation (3) adds it after k0k_0. These gaps affect the main inferential contributions and require more than cosmetic corrections.

  2. The calibration results do not support the claimed reliable and robust inference. In Table 1, the mean test’s rejection rate at nominal level 0.050.05 is only 0.00170.0017 for n=10000n=10000, and the reported values do not demonstrate convergence toward the nominal level. More seriously, Table 3 reports variance-test rejection rates of 0.1390.139, 0.2560.256, and 0.4360.436 under t5t_5 noise for n=150,450,2000n=150,450,2000, respectively, despite there being no change-point and the nominal level being 0.050.05. The appendix acknowledges this failure, but the main text still broadly claims robustness to heavy-tailed noise. Accurate localization conditional on detecting a change does not resolve failure to control false positives. These results materially limit the method’s practical reliability.

Questions for authors
  1. Can the authors provide the missing uniform positive separation argument for mean change-point localization in Appendix C.3? How does it establish the claimed rate for the actual estimator, whose smooth functions are fitted at candidate locations?

  2. Which transformed sequence is used in the implementation: Equation (3) or the construction in Appendix C.5? Can the authors justify the null distribution for the candidate-dependent fitted-residual statistic, including the effects of estimating the trends and variance parameters from the same data?

  3. What explains the severe conservativeness in Table 1 and the increasing false-positive rates under t5t_5 noise in Table 3? Were the reported power and localization results obtained using the same critical values and complete iterative procedure? Please clarify which robustness claims remain supported by these results.

Reviewer_2dHGOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies offline change detection when a smoothly evolving mean may contain one discontinuity and the variance may change at a different location. It proposes an alternating procedure combining weighted smoothing splines with mean- and variance-change tests, and claims estimation rates and asymptotic null distributions. Simulations and real data examples illustrate the method.

Critical strengths
  • Separating smooth evolution, mean shifts, and variance changes is practically meaningful.
  • Allowing different mean and variance change locations and estimating interpretable parameters are useful features.
Critical weaknesses

1. Theory gaps

Algorithm–proof mismatch. Equation (3) constructs variance adjusted sequence from fitted residuals and an estimated jump. Appendix C.5 instead analyzes one independent Gaussian sequence without establishing the necessary approximation.

A counterexample to the displayed procedure. Consider independent N(0,1)N(0,1) observations with zero mean, so there is neither a mean change nor a variance change. Supply the known unit variance to Algorithm 2. Use the periodic first-order spline setting, with m=1m=1 and smoothing parameter n−2/3n^{-2/3}, giving bandwidth h=n−1/3h=n^{-1/3}. These choices satisfy Theorem 4.2's stated bandwidth conditions.

Let KkK_k denote the estimated jump between the fitted boundary values at candidate split kk. Because the fits include the unpenalized constant component described in Section 3.1, residuals sum to zero separately on each segment. Equation (3) therefore gives calibrated segment means equal to zero on the left and KkK_k on the right. The candidate RSS reduction is exactly

D(k)=[k(n−k)/n]Kk2.D(k) = [k(n-k)/n] K_k^2.

For even nn, consider the midpoint k=n/2k=n/2. The periodic first-order spline calculation gives each fitted boundary value a variance asymptotic to 1/(2nh)1/(2nh). Since the two fits use independent halves of the observations, the variance of Kn/2K_{n/2} is asymptotic to 1/(nh)1/(nh). Consequently, 4hD(n/2)4hD(n/2) converges in distribution to a chi-square random variable with one degree of freedom.

The midpoint score thus grows on the scale 1/h=n1/31/h=n^{1/3}, whereas the proposed critical value for Tn2T_n^2 grows only on the log log n scale. Since Tn2T_n^2 is at least D(n/2)D(n/2), the null rejection probability tends to one, contradicting the claimed calibration.

This objection concerns equations (2)–(3) with the stated unpenalized constant component. The authors should reconcile the example with the fitting space and implemented statistic.

Other gaps.

  • Initialization and mean localization: The initialization lemmas assume the true split rather than Algorithm 2's estimated split. The mean-localization argument in Appendix C.3 concludes from an order bound without establishing a positive contrast away from the true change.

  • The first-order spline corollary: For m=1m=1, substitution into Theorem 4.1 supplies only an Op(n−1/3)O_p(n^{-1/3}) upper bound for the squared mean-function error, rather than establishing Corollary 4.1's Op(n−2/3)O_p(n^{-2/3}). Appendix C.4 assumes the sharper mean rate when deriving root-n variance estimation. An additional argument is needed; this does not establish that the sharper rates are impossible.

  • Variance-test calibration: Theorem 4.3, the critical-value derivation, and the proof use incompatible powers of log⁡n\log n. Separately, making Appendix C.6's displayed remainder bound vanish for m=1m=1 would require n1/3hn^{1/3}h to diverge while n1/2hn^{1/2}h tends to zero, which is incompatible. This identifies a failure of the submitted bound, rather than disproving every possible corrected variance theorem.

2. The novelty needs sharper positioning

The alternating weighted-spline/residual-variance framework, initialization, GCV, and residual-stability stopping already appear in Gao et al. (2019). The principal extension is the mean-jump block interacting with a separate variance change. Its validity is therefore central to the contribution.

Task-matched comparisons are needed: Gao's method is particularly relevant to variance-only experiments, while Wu and Zhou (2024) provides an aligned mean-jump competitor under changing noise dynamics.

The results concern one offline change per parameter under independent Gaussian errors. Estimating multiple changes would generally be more useful for retrospective analysis of longer series, while non-Gaussian and temporally dependent observations would better match many motivating applications.

Questions for authors
  1. Does the implementation exactly follow equations (2)–(3), including the unpenalized constant component? How is the Gaussian null example compatible with Theorem 4.2?
  2. What corrected assumptions and arguments establish the initialization, localization, m=1m=1 estimation rates, and variance-test calibration?
  3. Can the authors clarify the signal-strength scope of the localization results and provide unconditional performance, task-matched comparisons, and runtime measurements?
Reviewer_bbLyOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper alternates smoothing-spline estimation with tests for mean and variance changes. It considers smooth trends with abrupt changes at potentially different locations, and provides asymptotic results, simulations, and a battery-production example.

Critical strengths

Separating abrupt mean changes from variance changes in the presence of a smooth trend is a relevant statistical problem. Joint estimation could be useful when analyzing either component alone gives misleading results.

Critical weaknesses
  1. I cannot reconcile the test definitions with their proofs and calibration. The mean-test algorithm constructs a different transformed sequence for each candidate mean-change location. Appendix C.5 instead adds the estimated jump after the variance-change location and then analyzes a single transformed sequence. The variance test has a separate inconsistency: Theorem 4.3 and Appendix B.2 use different logarithmic scaling factors for the same statistic. These discrepancies directly affect which test the theory actually justifies. The mean-test proof also needs to account explicitly for estimated spline residuals and plugged-in variance estimates.

  2. The reported evidence does not establish validity of the final iterative procedure. Appendix C.3 propagates consistency between iterations, but replaces a justification for the data-dependent stopping rule with the observation that the algorithm usually converges quickly. The size results also need explanation: Appendix D.2 reports an empirical rejection rate of 0.0017 at nominal 0.05 even with 10,000 observations. That is far from nominal, and the table does not substantiate the stated convergence toward the target level.

Overall, I recommend rejection. The central statistical guarantees cannot currently be reconciled with the proposed procedure and calibration formulas.

Questions for authors

Which transformations and normalization are implemented? Can you reconcile the algorithm, theorem statements, and critical values, and explain whether the guarantees cover the actual stopping rule? Are the reported size values correct? These are clarification requests that could materially change my assessment.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.