Pith. sign in

REVIEW 2 major objections 4 minor 74 references

Ridge-regularized Hotelling CUSUM statistics for functional time series converge to a non-Gaussian weighted sum of squared Brownian bridges, which supplies direct, bootstrap-free calibration for detecting and dating mean changes.

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 ↗

T0 review · deepseek-v4-flash

2026-08-01 19:23 UTC pith:CJYJ7423

load-bearing objection A substantive, carefully proved functional-data change-point paper whose central weighted-bridge limit holds under the stated assumptions; the main caveats are the violated grid-resolution condition in the Niño application and the gap between the WBS theory and its implemented threshold. the 2 major comments →

arxiv 2607.16976 v1 pith:CJYJ7423 submitted 2026-07-18 stat.ME

Adaptive Ridge-Regularized Hotelling Change-Point Tests for Functional Data

Ping Zhao , Long Feng This is my paper
classification stat.ME MSC 62M1062H1560F1762G10
keywords functional time serieschange pointridge regularizationlong-run covariancewild binary segmentationweighted Brownian bridgedifference-based estimatorHotelling test
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.

This paper sets out to show that mean changes in weakly dependent functional time series can be detected and located by a ridge-regularized Hotelling framework that needs no bootstrap and no pre-selected principal components. It expands each observed curve in a growing Fourier basis, estimates the long-run covariance with an edge-corrected difference-based estimator, and standardizes ridge CUSUM paths with empirical spectral weights. The central claim is that, under the null, the centered statistics converge jointly to a non-Gaussian weighted sum of squared Brownian bridges, and that critical values for fixed ridges and for a Cauchy-transformed aggregate can be read directly from that limit. The paper further derives an explicit local-power formula showing that the power-optimal ridge depends on the unknown direction of the change, which motivates aggregating over a ridge grid. If the theory holds, the resulting tests have asymptotically controlled size, are consistent for fixed alternatives, and locate one or multiple changes consistently under weak dependence and non-Gaussian errors.

Core claim

The central discovery is the weighted squared-Brownian-bridge limit for the centered ridge CUSUM family: D_a(t)=Σ_j w_j(a){B_j²(t)/(t(1−t))−1}/{2A_2(a)}^{1/2}, with ridge weights w_j(a)=λ_j/(λ_j+a·tr(Ω)) determined by the trace-class spectrum of the long-run covariance operator. Because the covariance is trace class, low-frequency directions retain non-vanishing weight, so the limit is non-Gaussian, in contrast to proportional random-matrix regimes. The paper proves this null limit jointly over a finite ridge grid, shows that the same common bridge paths calibrate both fixed-ridge and Cauchy-aggregated statistics, and derives local-alternative limits, an explicit power formula, an oracle rid

What carries the argument

The key object is the weighted squared-Brownian-bridge limit process D_a(t) and its finite-grid joint version, which the paper uses directly for calibration. This is a centered and standardized weighted sum of independent squared Brownian bridges, where the weights come from the long-run covariance spectrum and the ridge parameter. The bridge paths are shared across all ridge values, preserving cross-ridge dependence so that a Cauchy transform of marginal p-values is calibrated by the joint law. Supporting it is the edge-corrected difference-based long-run covariance estimator, whose positive spectral part supplies the empirical weights and the ridge-inverted resolvent.

Load-bearing premise

The load-bearing premise is that the numerical Fourier scores computed on the observation grid approximate the true functional projections closely enough that χ_{p,N}=p^{5/2}/(N−1)^2→0; when the grid is too coarse for the basis dimension, the proof of the weighted-bridge limit does not go through.

What would settle it

Compute the discrepancy between the numerical Fourier operator J_{p,N} and the exact projection Π_p on C² functions at the resolution used in the monthly sea-surface-temperature application (p=9, N=12), where χ≈2.0. If the discrepancy is not small, simulate n null functional curves at that resolution and compare the empirical 95th percentile of the ridge CUSUM statistic with the weighted squared-Brownian-bridge critical value; a systematic size bias growing with n would show the resolution condition is necessary.

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

If this is right

  • No bootstrap: critical values come from simulating an explicit weighted squared-Brownian-bridge law, so the test is feasible for long functional sequences.
  • Weak dependence and non-Gaussian errors are covered: only Bernoulli-shift approximability and finite ν>4 moments are required, not Gaussianity or a linear-process model.
  • Every fixed ridge on the grid is consistent if any one ridge sees n·q_a(d_n)→∞; the Cauchy aggregate inherits this consistency.
  • The explicit local-power formula identifies the oracle ridge, which depends on the unknown change direction, explaining why a ridge grid rather than a single ridge is used.
  • For multiple changes, wild binary segmentation with a deterministic diverging threshold recovers the true number of changes with probability tending to one and estimates their locations to o_P(Δ_n) under proportional spacing.

Where Pith is reading between the lines

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

  • If the grid-resolution condition fails—for example, the paper's monthly sea-surface-temperature application uses p=9 basis functions on an N=12 grid, where χ≈2—the numerical-score approximation is outside the stated asymptotic regime; a direct finite-sample check would compare empirical null quantiles of T_{n,a} with the weighted-bridge critical values on such coarse grids.
  • The ridge-grid and Cauchy-aggregation idea transfers naturally to other function-space statistics, such as changes in covariance operators or eigensystems, where the same spectral-decay problem appears.
  • The localization rule—pick the ridge with smallest marginal p-value—could be replaced by data-driven ridge selection or location averaging; the paper proves uniform location consistency but does not claim optimal selection.
  • For sparsely or irregularly observed curves, the numerical-Fourier route would need joint treatment of reconstruction and measurement error, as the paper itself notes, suggesting a testable extension of the same difference-based estimator.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper develops a unified ridge-regularized Hotelling framework for detecting and dating mean changes in weakly dependent functional time series. Functional curves are mapped to growing-dimensional Fourier scores via numerical integration on a common grid; an edge-corrected difference-based estimator is used for the long-run covariance; ridge CUSUM statistics are calibrated directly against an explicit weighted squared-Brownian-bridge limit; a Cauchy transform aggregates evidence over a ridge grid. For multiple changes, the same global covariance estimate is reused in local max-ridge statistics within wild binary segmentation followed by local refinement. The main theorems establish a joint non-Gaussian weighted-bridge null limit (Theorem 3.1), plug-in calibration consistency (Theorem 3.2), local alternative limits and an oracle-ridge formula (Theorems 3.3–3.5), single-change localization (Theorem 3.6), and WBS exact recovery with localization (Theorem 4.1). The proofs are detailed and organized in appendices.

Significance. If correct, the paper provides a practically useful alternative to bootstrap-based functional change-point tests: the reference law is explicit, no resampling of curves is needed, weak dependence and non-Gaussian errors are accommodated, and the multiple-change procedure has consistency guarantees. The main theoretical contributions — the non-Gaussian weighted-bridge limit under spectral decay, the contamination bounds for difference-based covariance estimation under a mean shift, and the WBS recovery proof with a globally estimated long-run covariance — are substantial. A particular strength is the transparent treatment of numerical basis integration: the discrepancy from ideal projections is stated as an explicit condition rather than silently assumed away. The main weakness is that this same condition is violated in one of the paper’s two empirical applications, so the asymptotic claims do not cover that example as presented.

major comments (2)
  1. [Section 6.2, Assumption 3.2, Lemma A.3] The Niño 1+2 application uses N=12 monthly grid points and p=9 Fourier coefficients. Assumption 3.2 requires χ_{p,N}=p^{5/2}/(N-1)^2 → 0; here χ_{p,N}=243/121≈2.0. Lemma A.3 bounds the numerical-score discrepancy by Cχ_{p,N}∥f∥_{C^2}, so the implemented trapezoidal Fourier scores are not guaranteed to be close to the ideal L2 projections in this application. Since Lemma A.4 and the proof of Theorem 3.1 use this bound to pass from the numerical-score CUSUM to the ideal projection process, the weighted-bridge null limit and the local-alternative limits are not justified for the statistic actually computed on this data set. The p-value 0.0365 and the 1981/82 boundary are therefore outside the asymptotic regime of Theorems 3.1–3.5. The paper is explicit about the condition, so this is not an internal inconsistency; however, it is load-bearing for the empirical claim and for the advertised ro
  2. [Section 4.1 / Assumption 4.1] The multiple-change theory assumes a fixed number J_cp of changes and proportional spacing Δ_n ≥ c_Δ n. This is stated as an assumption, and the conclusion section lists diverging J_cp and shrinking spacing as future work. I do not regard this as an error, but it should be kept visible when advertising “consistent recovery” — the current guarantee is for well-separated, proportionally spaced breaks only, which is a strong structural condition for functional data.
minor comments (4)
  1. [Equation (2.12)] In the mid-rank construction, the expression 1/2 + Σ 1{...} / (N_sim+1) is not a standard empirical cdf; consider clarifying that this is intentionally used to keep values away from 0 and 1, and explain the factor 1/2 in terms of continuous averaging.
  2. [Section 6.2, around Table 7] Please state the value of χ_{p,N} explicitly when reporting N=12, p=9, and relate it to Assumption 3.2. The current text reports reconstruction error 0.065 but does not connect it to the formal condition, which is misleading because the reconstruction error and the quadrature error are different quantities.
  3. [Section 3.1, Eq. (3.8)] The covariance formula is correct, but it would help to note that the sum is finite because |w_j(a)w_j(a')| ≤ 1 and A_2(a), A_2(a') > 0, so no additional summability assumption is needed.
  4. [Section 5.2, Table 1] RHT is mildly liberal in several cells (e.g., 0.083 for FAR setting 2, n=200, p=41). With 1,000 replications this is about 4–5 standard errors above 0.05; consider a brief comment on whether this reflects finite-sample bias in the bridge approximation or Monte Carlo calibration noise.

Circularity Check

0 steps flagged

No significant circularity: the null limit and plug-in calibration are derived from stated assumptions, not from the fitted answers.

full rationale

The paper's central derivation is self-contained. Theorem 3.1 derives the weighted squared-Brownian-bridge null limit from Assumptions 3.1–3.4 via the functional FCLT (Lemma A.2), numerical projection error bounds (Lemma A.3), the difference-based long-run covariance rate (Lemma A.8), and spectral-weight consistency (Lemma A.10). The limiting process in (3.5) is defined from the population long-run covariance spectrum, not from the estimated weights used in the statistic. The calibration in (2.11) does plug estimated spectral weights into the reference law, but Theorem 3.2 proves that the plug-in reference converges to the population law uniformly, so this is legitimate plug-in calibration rather than fitting the answer or renaming the statistic. The local-power formulas and the WBS consistency theorem are also obtained by derivation from the stated assumptions. The self-citations to Zhao et al. (2026) and Liu et al. (2026) are contextual references; the paper independently proves the Cauchy-aggregation and difference-based covariance results it actually uses, so they are not load-bearing. The reviewer's Assumption 3.2 concern for the Nino 1+2 application is an applicability/robustness issue outside the stated asymptotic regime, not a circularity in the derivation chain.

Axiom & Free-Parameter Ledger

7 free parameters · 8 axioms · 0 invented entities

No new physical or mathematical entities are postulated; the weighted-Brownian-bridge process is a derived limit, not an invented object. The paper's contribution is a method plus theory. Free parameters are all pre-specified tuning choices (ridge grid, filter, bandwidths, kernel, trimming, Cauchy weights, WBS parameters), none fitted to data; the authors disclose their heuristic character where relevant.

free parameters (7)
  • ridge grid G = {0.1, 0.2, 0.4, 0.8} in simulations; theory permits any fixed finite grid in (0,∞)
    Section 2.3 and 5.1. Fixed a priori, not fitted. Results depend on it through the reference weights; the theory explicitly excludes a_n→0 or a_n→∞, bounding the adaptivity claim to a user-chosen grid.
  • difference filter (m; c_0,...,c_m) = m=3; (0.1942, 0.2809, 0.3832, -0.8582)
    Section 5.1. Chosen by hand to satisfy (2.3); the asymptotic rates only require fixed m.
  • bandwidth ℓ and spacing h = ℓ=3, h=6 (IID); ℓ=5, h=10 (FAR); theoretical guidance ℓ≍n^{1/4}, h=2ℓ
    Section 5.1. Determines r_n=(ℓ/N_D)^{1/2}+ℓ^{-β_K}; hand-chosen per dependence structure.
  • kernel K = K(x)=(1-x²)_+
    Sections 2.2 and 5.1. Quadratic kernel with β_K=2; not positive semidefinite in finite samples, requiring the positive-part projection (2.7).
  • trimming ϵ and scan set K_{n,ϵ} = ϵ=0.1
    Section 2.3. Standard trim to avoid edge singularities in t(1-t).
  • Cauchy weights ϖ_v = equal weights 1/L_G
    Section 2.4. User-specified; theory allows any positive weights summing to 1.
  • WBS interval count M_n and threshold Λ_n = M=500 in simulations; Λ_n via Gaussian-reference quantile in practice; deterministic rate n^{γ_W}, γ_W∈(1/2+1/ν,1) in Th
    Sections 4.2-4.3 and 5.5. The theorem needs a deterministic diverging threshold satisfying (4.3); the implemented threshold is a conditional reference quantile explicitly not covered by Theorem 4.1.
axioms (8)
  • domain assumption Assumption 3.1: e_i = G(η_i, η_{i-1},...) with L^ν M-approximability, ν>4, Σ(1+M)^{β_dep} d_ν(M)<∞, H^ς, ς>5/2
    The FCLT (Lemma A.2), fourth-order summability (Lemma A.1), and DB-estimator rates (Lemma A.7) all rest on it. This is a standard ARS-type condition but excludes heavy-tailed errors with fewer than four moments.
  • domain assumption Assumption 3.2: μ, δ_n ∈ C²[0,1] with uniform C² bounds; p^{5/2}/(N-1)² → 0
    Controls the numerical Fourier integration error (Lemma A.3). Violated in the Niño 1+2 application (N=12, p=9 gives χ≈2.0).
  • domain assumption Assumption 3.3: fixed m with (2.3); even kernel K(0)=1, |1-K(x)|≤C|x|^{β_K}; ℓ→∞, 2ℓ≤h≤C_hℓ, ℓ+mh=o(n)
    Standard bandwidth growth for difference-based estimation.
  • domain assumption Assumption 3.4: tr(Ω)>0 and √p·r_n → 0 (under ℓ≍n^{1/4} this permits p=o(n^{3/4}))
    Non-degeneracy plus the rate needed for trace-norm control of the positive-part estimator.
  • domain assumption Assumption 4.1: fixed J_cp, proportional spacing Δ_n ≥ c_Δ n, δ_min > 0
    Enables isolating intervals in the WBS proof; the paper itself identifies diverging J_cp and shrinking spacing as future work.
  • domain assumption CR_α: G_C continuous at its (1-α) quantile with strict crossing
    Identifies a unique nonrandomized limiting critical value; holds automatically for singleton ridge grids; needed only for exact level of the joint Cauchy statistic.
  • standard math Functional Gaussian coupling of Berkes-Philipp-Jirak (Theorem A.1 of Aue et al. 2017, restated as Lemma B.5)
    Used to obtain D([0,1],H) invariance under the ARS approximation condition.
  • standard math Pinelis martingale maximal inequality; Massart DKW bound; Hoffman-Wielandt; Lidskii-Mirsky-Wielandt; PSD metric projection property
    Used in Lemmas B.4, A.13, A.10, and (B.1).

pith-pipeline@v1.3.0-alltime-deepseek · 64351 in / 25356 out tokens · 236947 ms · 2026-08-01T19:23:48.419258+00:00 · methodology

0 comments
read the original abstract

We propose a unified ridge-regularized Hotelling framework for detecting and locating mean changes in functional time series. A growing basis expansion converts the functional observations into high-dimensional score vectors. Their long-run covariance is estimated by an edge-corrected difference-based procedure. Ridge regularization stabilizes inference under spectral decay. An explicit local-power formula shows that the power-maximizing ridge depends on the unknown spectral orientation of the change. We therefore combine a family of ridge CUSUM statistics by a Cauchy transform and calibrate the aggregate directly from their joint weighted-bridge limit. For multiple changes, we embed local maximum-ridge statistics in a wild binary segmentation procedure, followed by local refinement. Under mild conditions, we establish the validity, local power, consistency, and localization properties of the proposed tests. In the multiple-change setting, the procedure consistently recovers the number of changes and uniformly estimates their locations. The framework accommodates weak dependence and non-Gaussian functional errors. Simulations and two empirical applications demonstrate the favorable finite-sample performance of the proposed methods.

Figures

Figures reproduced from arXiv: 2607.16976 by Long Feng, Ping Zhao.

Figure 1
Figure 1. Figure 1: Empirical power under IID errors. The rows correspond to the finite-rank, fast-decay, and slow-decay innovation spectra, and the columns correspond to ssig = 5 and ssig = 20 randomly selected signal directions. Each point is based on 100 paired replications [PITH_FULL_IMAGE:figures/full_fig_p023_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Empirical power under FAR errors with ∥Ψ∥op = 0.5. The panel arrangement is the same as in [PITH_FULL_IMAGE:figures/full_fig_p024_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sydney annual minimum-temperature curves around the RHT change point. The left panel shows the phase mean curves before and after the retained WBS change. The right panel shows their daywise difference and its 21-term Fourier representation. The Fourier decomposition in [PITH_FULL_IMAGE:figures/full_fig_p027_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Fourier decomposition of the Sydney phase means. The upper panel shows the coefficients of the two fitted phase means. The lower panel shows the signed coefficients of their difference [PITH_FULL_IMAGE:figures/full_fig_p028_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: NOAA ERSSTv6 Niño 1+2 analysis. The left panel shows the annual monthly curves and the two fitted segment means. The middle panel is the RHT localization path selected by the smallest marginal ridge p-value. The right panel shows the sample standard deviations of the retained Fourier scores. Residual dependence is assessed after subtracting the two fitted segment means. We exclude lagged pairs crossing the… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

74 extracted references

  1. [1]

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

    Aue, Alexander and Rice, Gregory and S\"onmez, Ozan , title =. Journal of the Royal Statistical Society Series B: Statistical Methodology , year =

  2. [2]

    2017 , howpublished=

    Aue, Alexander and Rice, Gregory and S\"onmez, Ozan , title =. 2017 , howpublished=

    2017

  3. [3]

    Detecting changes in the mean of functional observations , journal =

    Berkes, Istv. Detecting changes in the mean of functional observations , journal =. 2009 , volume =

    2009

  4. [4]

    Journal of Business & Economic Statistics , year =

    Chan, Kin Wai , title =. Journal of Business & Economic Statistics , year =

  5. [5]

    The Annals of Statistics , year =

    Chan, Kin Wai , title =. The Annals of Statistics , year =

  6. [6]

    and Paul, Debashis and Prentice, Ross L

    Chen, Lin S. and Paul, Debashis and Prentice, Ross L. and Wang, Pei , title =. Journal of the American Statistical Association , year =

  7. [7]

    , title =

    Fremdt, Stefan and Horv\'ath, Lajos and Kokoszka, Piotr and Steinebach, Josef G. , title =. Journal of Multivariate Analysis , year =

  8. [8]

    The Annals of Statistics , year =

    H\"ormann, Siegfried and Kokoszka, Piotr , title =. The Annals of Statistics , year =

  9. [9]

    The Annals of Statistics , year =

    Li, Haoran and Aue, Alexander and Paul, Debashis and Peng, Jie and Wang, Pei , title =. The Annals of Statistics , year =

  10. [10]

    2026 , howpublished=

    Liu, Yanhong and Song, Fengyi and Feng, Long , title =. 2026 , howpublished=

    2026

  11. [11]

    Journal of the American Statistical Association , year =

    Liu, Yaowu and Xie, Jun , title =. Journal of the American Statistical Association , year =

  12. [12]

    and West, Kenneth D

    Newey, Whitney K. and West, Kenneth D. , title =. Econometrica , year =

  13. [13]

    and Tavakoli, Shahin , title =

    Panaretos, Victor M. and Tavakoli, Shahin , title =. The Annals of Statistics , year =

  14. [14]

    Journal of Time Series Analysis , year =

    Rice, Gregory and Shang, Han Lin , title =. Journal of Time Series Analysis , year =

  15. [15]

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

    Wang, Guanghui and Feng, Long , title =. Journal of the Royal Statistical Society Series B: Statistical Methodology , year =

  16. [16]

    2025 , howpublished=

    Wang, Xiaoyi and Liu, Jixuan and Feng, Long , title =. 2025 , howpublished=

    2025

  17. [17]

    2026 , howpublished=

    Zhao, Ping and Zhou, Le and Feng, Long , title =. 2026 , howpublished=

    2026

  18. [18]

    Billingsley, Patrick , title =

  19. [19]

    Bhatia, Rajendra , title =

  20. [20]

    Simon, Barry , title =

  21. [21]

    Andrews, Donald W. K. , title =. Econometrica , year =

  22. [22]

    , title =

    Higham, Nicholas J. , title =. Linear Algebra and its Applications , year =

  23. [23]

    The Annals of Probability , year =

    Massart, Pascal , title =. The Annals of Probability , year =

  24. [24]

    Statistics & Probability Letters , year =

    Jirak, Moritz , title =. Statistics & Probability Letters , year =

  25. [25]

    and Fournier, John J

    Adams, Robert A. and Fournier, John J. F. , title =

  26. [26]

    , title =

    Brillinger, David R. , title =

  27. [27]

    and van Zanten, Harry J

    van der Vaart, Aad W. and van Zanten, Harry J. , title =. Pushing the Limits of Contemporary Statistics: Contributions in Honor of Jayanta K. Ghosh , editor =

  28. [28]

    , title =

    Rhee, Wansoo T. , title =. Annales de l'Institut Henri Poincar. 1984 , volume =

    1984

  29. [29]

    , title =

    Bogachev, Vladimir I. , title =

  30. [30]

    and Silverman, Bernard W

    Ramsay, James O. and Silverman, Bernard W. , title =

  31. [31]

    Ferraty, Fr\'ed\'eric and Vieu, Philippe , title =

  32. [32]

    Horv\'ath, Lajos and Kokoszka, Piotr , title =

  33. [33]

    Journal of Multivariate Analysis , year =

    Aue, Alexander and Gabrys, Robertas and Horv\'ath, Lajos and Kokoszka, Piotr , title =. Journal of Multivariate Analysis , year =

  34. [34]

    Aston, John A. D. and Kirch, Claudia , title =. Journal of Multivariate Analysis , year =

  35. [35]

    , title =

    Zhang, Xianyang and Shao, Xiaofeng and Hayhoe, Katharine and Wuebbles, Donald J. , title =. Electronic Journal of Statistics , year =

  36. [36]

    2015 , howpublished=

    Torgovitski, Leonid , title =. 2015 , howpublished=

    2015

  37. [37]

    The Canadian Journal of Statistics , year =

    Sharipov, Olimjon and Tewes, Johannes and Wendler, Martin , title =. The Canadian Journal of Statistics , year =

  38. [38]

    Electronic Journal of Statistics , year =

    Bastian, Patrick , title =. Electronic Journal of Statistics , year =

  39. [39]

    The Annals of Applied Statistics , year =

    Chiou, Jeng-Min and Chen, Yu-Ting and Hsing, Tailen , title =. The Annals of Applied Statistics , year =

  40. [40]

    Statistics & Probability Letters , year =

    Rice, Gregory and Zhang, Chi , title =. Statistics & Probability Letters , year =

  41. [41]

    Journal of the American Statistical Association , year =

    Chen, Yu-Ting and Chiou, Jeng-Min and Huang, Tzee-Ming , title =. Journal of the American Statistical Association , year =

  42. [42]

    2026 , howpublished=

    Bai, Lujia and Hu, Qirui and Wu, Weichi , title =. 2026 , howpublished=

    2026

  43. [43]

    Dependent functional linear models with applications to monitoring structural change , journal =

    Aue, Alexander and H\"ormann, Siegfried and Horv\'ath, Lajos and Hu. Dependent functional linear models with applications to monitoring structural change , journal =. 2014 , volume =

    2014

  44. [44]

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

    Gromenko, Oleksandr and Kokoszka, Piotr and Reimherr, Matthew , title =. Journal of the Royal Statistical Society Series B: Statistical Methodology , year =

  45. [45]

    Stoehr, Christina and Aston, John A. D. and Kirch, Claudia , title =. Econometrics and Statistics , year =

  46. [46]

    Electronic Journal of Statistics , year =

    Dette, Holger and Kutta, Tim , title =. Electronic Journal of Statistics , year =

  47. [47]

    Change point analysis of covariance functions: A weighted cumulative sum approach , journal =

    Horv. Change point analysis of covariance functions: A weighted cumulative sum approach , journal =. 2022 , volume =

    2022

  48. [48]

    The Annals of Statistics , year =

    Jirak, Moritz , title =. The Annals of Statistics , year =

  49. [49]

    , title =

    Wang, Tengyao and Samworth, Richard J. , title =. Journal of the Royal Statistical Society Series B: Statistical Methodology , year =

  50. [50]

    The Annals of Statistics , year =

    Enikeeva, Farida and Harchaoui, Zaid , title =. The Annals of Statistics , year =

  51. [51]

    , title =

    Liu, Haoyang and Gao, Chao and Samworth, Richard J. , title =. The Annals of Statistics , year =

  52. [52]

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

    Yu, Mengjia and Chen, Xiaohui , title =. Journal of the Royal Statistical Society Series B: Statistical Methodology , year =

  53. [53]

    Journal of the American Statistical Association , year =

    Zhang, Yangfan and Wang, Runmin and Shao, Xiaofeng , title =. Journal of the American Statistical Association , year =

  54. [54]

    The Annals of Statistics , year =

    Wang, Runmin and Zhu, Changbo and Volgushev, Stanislav and Shao, Xiaofeng , title =. The Annals of Statistics , year =

  55. [55]

    2026 , howpublished=

    Li, Haoran and Xu, Haotian , title =. 2026 , howpublished=

    2026

  56. [56]

    Statistical Papers , year =

    Wegner, Lea and Wendler, Martin , title =. Statistical Papers , year =

  57. [57]

    Cooper and Horv\'ath, Lajos and Trapani, Lorenzo , title =

    Boniece, B. Cooper and Horv\'ath, Lajos and Trapani, Lorenzo , title =. Journal of Econometrics , year =

  58. [58]

    Journal of Time Series Analysis , year =

    Horv\'ath, Lajos and Rice, Gregory and VanderDoes, Jeremy , title =. Journal of Time Series Analysis , year =

  59. [59]

    Derek , title =

    Harris, Trevor and Li, Bo and Tucker, J. Derek , title =. Environmetrics , year =

  60. [60]

    The Annals of Applied Statistics , year =

    Bastian, Patrick and Basu, Rupsa and Dette, Holger , title =. The Annals of Applied Statistics , year =

  61. [61]

    and Ombao, Hernando , title =

    Jiao, Shuhao and Frostig, Ron D. and Ombao, Hernando , title =. Scandinavian Journal of Statistics , year =

  62. [62]

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

    Cho, Haeran and Fryzlewicz, Piotr , title =. Journal of the Royal Statistical Society Series B: Statistical Methodology , year =

  63. [63]

    Electronic Journal of Statistics , year =

    Cho, Haeran , title =. Electronic Journal of Statistics , year =

  64. [64]

    The Annals of Probability , year =

    Pinelis, Iosif , title =. The Annals of Probability , year =

  65. [65]

    The Annals of Statistics , year =

    Fryzlewicz, Piotr , title =. The Annals of Statistics , year =

  66. [66]

    Functional data analysis for sparse longitudinal data , journal =

    Yao, Fang and M. Functional data analysis for sparse longitudinal data , journal =. 2005 , volume =

    2005

  67. [67]

    Functional linear regression analysis for longitudinal data , journal =

    Yao, Fang and M. Functional linear regression analysis for longitudinal data , journal =. 2005 , volume =

    2005

  68. [68]

    Functional variance processes , journal =

    M. Functional variance processes , journal =. 2006 , volume =

    2006

  69. [69]

    Climate Data Online , year =

  70. [70]

    Site summary: Sydney (Observatory Hill), station 066062 , year =

  71. [71]

    2026 , howpublished=

    Australian Climate Observations Reference Network--Surface Air Temperature (. 2026 , howpublished=

    2026

  72. [72]

    State of the Climate 2024 , year =

    2024

  73. [73]

    2026 , howpublished=

    Ni. 2026 , howpublished=

    2026

  74. [74]

    2026 , howpublished=

    Extended Reconstructed Sea Surface Temperature (. 2026 , howpublished=

    2026