{"id":"0aed6721-24fe-4947-8672-dd9dfba3da59","arxiv_id":"2502.08430","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A bootstrap method produces simultaneous confidence bands for mean functional curves in each segment between relevant change points in a stationary functional time series.","lead":"This paper builds statistical confidence bands for the average joint-angle curve of a runner during each fatigue phase, after a separate algorithm has found where fatigue states start and end. The bands show how a runner's range of motion changes as they tire, which could help spot risky movement patterns.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's coverage claim hinges on an unstated rate for the change-point and relevant-set estimators; if the BINSEG estimators converge slower than o_P(n^{-1/2}), the proof's substitution step is invalid and the bands can undercover.","rationale":"Reader's verdict is CONDITIONAL, and I find no reason to change it. The method is a reasonable extension of Dette et al. (2020), and the cited estimators plausibly have the required rates, but the proof of Theorem 1 does not state the external rate and separation conditions that make the substitution steps valid. My most specific concern is Step 1: without an o_P(n^{1/2}) bound on the change-point index error, the estimated segment mean differs from the true segment mean by a non-negligible term under the bootstrap scaling. A second, closely related concern is Step 2: the no-exact-Delta condition is not a sufficient separation condition in a triangular-array model, since a jump at Delta + n^{-1/2} satisfies it while preventing consistent relevant-set estimation. Both concerns are testable by writing out the cited theorems; if the rates are as expected, the proof sketch can be completed and the result is sound. This supports the original CONDITIONAL verdict rather than a full rejection.","tokens_in":7031,"tokens_out":15537,"duration_ms":176117,"concrete_test":"Extract the exact statements of Theorems 4.1 and 4.2 of Bastian et al. (2024) and Conditions (A1)-(A4) of Dette and Kokot (2022). Verify that Theorem 4.1 implies max_i n|s_hat_i - s_i| = O_P(1), which gives the needed o_P(n^{1/2}) index error, and that Theorem 4.2 gives I_hat = I with probability 1-o(1) under the stated no-jump-exactly-Delta condition without a fixed positive gap. If either implication fails, construct the triangular-array example with jump norm Delta + n^{-1/2}; if the empirical coverage of the bands in (11) converges to 1-alpha-beta rather than 1-alpha, the equality claim (13) is false as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is Theorem 1, equations (12)-(13), and its proof is a three-step sketch whose load-bearing step is Step 1: estimated change points are replaced by true change points in the definitions of nhat_i and muhat_i. For the replacement to be asymptotically negligible, the index error must be o_P(n^{1/2}) (equivalently, the fractional error s_hat_i - s_i must be o_P(n^{-1/2})); otherwise the misallocated boundary block contributes a bias of order delta * h / sqrt(n) to sqrt(n_i)(muhat_i - mu_i), which is not negligible. The paper cites Theorem 4.1 of Bastian et al. (2024) but does not state the rate it provides, nor the Conditions (A1)-(A4) from Dette and Kokot (2022) under which it holds. Step 2 has the same structure: equality in (13) requires I_hat = I with probability 1-o(1), but the only condition stated is the absence of jumps exactly equal to Delta. In the triangular-array model of (1)-(2), jump sizes may depend on n; a jump of norm Delta + n^{-1/2} satisfies 'no jump is exactly Delta' for every n, yet the relevant-set estimator cannot stabilize, so P(I_hat = I) does not tend to 1 and (13) is not justified. Thus the theorem rests on unstated rate and separation conditions that are essential to the proof.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes uniform confidence bands for segment means of a functional time series after multiple change point detection, targeting only segments whose adjacent mean jumps exceed a threshold Δ. The construction combines three ingredients: change point estimates and an estimated relevant-change set from Bastian et al. (2024), a long-run variance estimator of functional type, and a multiplier block bootstrap. The central theoretical statement, Theorem 1, claims that the bands defined in (11) have asymptotic simultaneous coverage at least 1−α−β, and exactly 1−α when no jump size equals Δ. The proof is only a three-step sketch that invokes external results from Bastian et al. (2024), Dette and Kokot (2022), and Dette et al. (2020). An application to knee-angle data of runners under fatigue illustrates the method. The paper is a short note: the motivation is clear and the idea is plausible, but the theorem as stated and proved is not currently verifiable from the manuscript.","tokens_in":7349,"tokens_out":13421,"duration_ms":137621,"significance":"If the claimed coverage result is correct, the paper would provide a practically useful post-hoc inference tool for biomechanical fatigue analysis, extending the change point methodology of Bastian et al. (2024) to simultaneous confidence statements about segment means. The combination of change point selection with bootstrap inference is of genuine interest, and the real-data examples demonstrate a relevant application. However, the significance is currently limited by the absence of a complete, self-contained proof: the central theorem depends on unstated conditions and unstated localization rates, and there is no simulation study or finite-sample check. The application is illustrative rather than a validation of the coverage property.","major_comments":[{"comment":"The proof replaces estimated change points by true change points in the definitions of n_hat_i and mu_hat_i, but no localization rate is stated. For the replacement to be asymptotically negligible, one needs max_i |s_hat_i − s_i| = o_P(n^{−1/2}); otherwise the misallocated boundary block contributes a bias of order sqrt(n) |s_hat_i − s_i| times the jump size to sqrt(n_hat_i)(mu_hat_i − mu_i). The citation to Theorem 4.1 of Bastian et al. (2024) is insufficient unless that theorem is stated together with the rate it provides and the conditions under which it holds. This is load-bearing because Step 1 is the first step of the proof of (12) and (13).","section":"§2, proof of Theorem 1, Step 1"},{"comment":"The condition that ||mu_{i+1} − mu_i||_∞ is not equal to Delta for all i does not suffice for the equality statement (13) under the triangular-array model (1)–(2). Because the segment means may depend on n, a jump of size Delta + n^{−1/2} is not equal to Delta for any n, yet the relevant-change estimator cannot stabilize, so P(I_hat = I) need not tend to 1. The proof's Step 2 therefore needs a separation condition, for example all relevant jumps at least Delta + ε_n and all non-relevant jumps at most Delta − ε_n with ε_n large enough for consistent detection, or an explicit assumption that the mu_i are fixed and hence a positive gap exists.","section":"§2, Theorem 1, condition for (13)"},{"comment":"The paper repeatedly invokes Conditions (A1)–(A4) of Dette and Kokot (2022) without stating them, and the proof of Lemma 1 dismisses the effect of estimated means by saying that the general result follows by straightforward approximation arguments. Since Theorem 1 and Lemma 1 are the theoretical core, the manuscript should either reproduce these conditions or state the precise assumptions on the noise process epsilon, including the mixing and moment conditions and the sense in which the process is valued in C[0,1]. The approximation step in Lemma 1 also requires the same change point localization rates as Step 1 and should be spelled out.","section":"§2, Theorem 1 and Lemma 1"},{"comment":"There is an internal inconsistency in the critical value used for the bands. The statistic T_n in (6) and its bootstrap version T*_n in (10) are sup-norm statistics, so the nominal (1−α) simultaneous bands should use the (1−α) quantile of T_n. Equations (7) and (11), however, use q_{1−α/2} (and (11) uses the bootstrap version q_hat*_{1−α/2}), while Algorithm 1 line 11 computes q_hat*_{1−α}. With the quantile as written in (11), the equality in (13) does not follow; the coverage would be at a different level or only an inequality. Please reconcile the definition of the band and the quantile.","section":"§2, eqs. (6), (7), (11) and Algorithm 1"},{"comment":"The event in (12) and (13) is written as the intersection over all i in I of events involving mu_hat_i^pm, but (11) defines mu_hat_i^pm only for i in I_hat. If I_hat differs from I, the probability statement is not formally well-defined. The bands should be defined for all candidate indices (with arbitrary values outside I_hat) or the intersection should be over I ∩ I_hat, with the negligible difference handled explicitly in the proof.","section":"§2, Theorem 1 statement"}],"minor_comments":[{"comment":"In line 1, the set of estimated change points is written as {s_hat_1, ..., s_hat_k_hat}, but the estimated number of change points is denoted m_hat elsewhere; the notation should be consistent.","section":"Algorithm 1"},{"comment":"The sum in line 8 runs to floor(n s_hat_{i+1}), whereas (8) and (9) use floor(n s_hat_{i+1}) − 1; the algorithm and the formula should agree.","section":"Algorithm 1, line 8"},{"comment":"Equation (7) uses sqrt(n) in the denominator while the final bands in (11) use sqrt(n_hat_i); since (7) is not used in the theorem, either align it with (11) or remove it to avoid confusion.","section":"§2, eq. (7)"},{"comment":"There are several typographical errors: 'theses results' should be 'these results', 'chnage' should be 'change', and the sentence 'In this paper will explicate' is missing a subject.","section":"Throughout"},{"comment":"The set I includes 0 for notational convenience, but mu_0 is not otherwise defined before its use in the expression mu_1 − mu_0; please define mu_0 explicitly.","section":"§2, eq. (4)"},{"comment":"The paper contains no simulation study or finite-sample assessment of the coverage of the proposed bands; at least a small Monte Carlo experiment matching the model in (1)–(2) would substantially strengthen the practical claims.","section":"§3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads like an extended abstract of a work in progress. The core theorem is plausible, but the proof is only a sketch and several technical conditions are external to the paper. The quantile inconsistency in (11) versus Algorithm 1 is especially important to fix, as it directly affects the claimed coverage level. I would support a major revision rather than rejection, provided the authors are willing to state the conditions and rates explicitly and add either a complete proof or a more detailed proof with the needed separation assumptions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something sensible: it takes the uniform-band machinery from Dette et al. (2020) for a single segment and extends it to the multiple-segment case after change point detection, with a relevance threshold. That is a legitimate incremental contribution, and it is clearly motivated by the biomechanics application. The writing is honest about leaning on earlier work, and the data examples are illustrative without being oversold.\n\nThe soft spot is exactly where the stress-test points. Theorem 1 is the whole paper, and its proof is a three-step sketch. Step 1 replaces estimated change points by true ones in the definitions of nhat_i and muhat_i. That is only valid if the change point estimators are accurate enough, e.g. fractional error o_P(n^{-1/2}); the paper cites Theorem 4.1 of Bastian et al. (2024) but does not state the rate it supplies, nor the Conditions (A1)-(A4) it needs. Without that rate, the substitution can leave a bias that does not vanish. Step 2 claims Ihat = I with probability 1-o(1), but the only condition stated is that no jump size equals Delta exactly. In the triangular-array model, jump sizes may depend on n, so a jump of size Delta + n^{-1/2} is never exactly Delta, yet the relevant-set estimator may not stabilize. Equality in (13) therefore does not follow from the stated assumptions. This is a real gap, not a manufactured one.\n\nI also share the reader's note that there are no simulations or code, which is a genuine omission for an inferential method. The consistency lemma for the long-run variance is also proved only up to a reference to a supplement, though that part is likely fine. The application uses a data-dependent Delta, which is reasonable in context but further obscures the separation condition.\n\nAll of this is addressable. The core idea is plausible and the paper is a clear, readable contribution for statisticians working on functional time series and change point analysis, and for biomechanists who want simultaneous bands. But as written, the coverage guarantee is not established. A serious referee should not desk-reject it; they should send it back for a completed proof, stated conditions, and a simulation study.","headline":"A useful but under-proved extension of Dette et al. (2020): the bootstrap bands are plausible, but the main theorem relies on unstated rates and a separation condition that the proof sketch does not supply.","tokens_in":7873,"tokens_out":2016,"would_cite":false,"duration_ms":24045,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G15","62G20","62M10","62R10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper constructs simultaneous uniform confidence bands for the mean curves of the segments between relevant change points in a functional time series, using a multiplier bootstrap that remains valid when the change locations and…","keywords":["uniform confidence bands","functional time series","change point detection","multiplier bootstrap","long run variance","fatigue biomechanics","range of motion","knee angle data"],"falsifier":"Simulate functional time series that satisfy the assumed conditions, with known change points and one jump size exactly equal to $\\Delta$, and compute empirical coverage of the bands over many replications; Theorem 1 predicts coverage at least $1-\\alpha-\\beta$ in that case and exactly $1-\\alpha$ when no jump equals $\\Delta$, so a systematic shortfall would refute the claim.","tokens_in":6820,"feed_emoji":"🏃","tokens_out":10721,"duration_ms":94667,"temperature":0.7,"pith_summary":"This paper extends post-hoc change point analysis for functional time series to confidence statements. It constructs uniform confidence bands for the segment mean curve of each phase where the change in the mean exceeds a threshold $\\Delta$, so that all relevant segments are covered simultaneously with asymptotic probability at least $1-\\alpha$ in the generic case. The motivation is biomechanical: locating fatigue states in a runner's lower-extremity joint angles and then describing the range of motion within each state. The bands are built from a multiplier bootstrap on block sums, after using existing estimators for change points and the relevant change set. If the asymptotic result holds, practitioners can compare rest, pre-fatigue and fatigued movement patterns with simultaneous statistical guarantees.","feed_headline":"Bootstrap bands cover every fatigue phase at once","feed_subtitle":"For knee-angle curves, one band construction covers rest, pre-fatigue, and fatigue segments simultaneously.","key_machinery":"The load-bearing mechanism is a multiplier bootstrap for the studentized statistic $T_n = \\max_{i\\in\\hat I} \\sqrt{\\hat n_i}\\, \\|(\\hat\\mu_i - \\mu_i)/\\hat\\sigma\\|_\\infty$, where $\\hat I$ is the estimated set of relevant change points. Bootstrap copies $\\hat\\mu_i^*$ are formed by multiplying block sums of centered residuals $Y_{n,j} = X_{n,j} - \\hat\\mu_i$ by independent standard normals, then normalizing by the long-run variance estimator $\\hat\\sigma^2$ of (15); the $(1-\\alpha)$ quantile of these copies sets the band width in (11). The construction inherits its validity from the preliminary estimators: the proof uses consistency of the change point and relevant-set estimators to replace estimated segments by true segments, and consistency of $\\hat\\sigma^2$ to replace $\\hat\\sigma$ by $\\sigma$. The block length $L$ and bandwidth $c$ are tuning parameters.","core_discovery":"Theorem 1 is the central claim: under Conditions (A1)--(A4) of Dette and Kokot (2022), a consistent long-run variance estimator, and consistent estimators of the change points and relevant set from Bastian et al. (2024), the bands in (11) satisfy $\\liminf_n P(\\cap_{i\\in I} \\{\\forall t: \\mu_i^-(t)\\le \\mu_i(t)\\le \\mu_i^+(t)\\}) \\ge 1-\\alpha-\\beta$, with equality to $1-\\alpha$ when no true jump size equals $\\Delta$. The bootstrap quantile of the studentized sup-statistic $T_n$, taken over all estimated relevant segments, therefore produces simultaneous bands rather than pointwise ones. The proof replaces estimated change points and estimated relevant indices by their true values, then uses weak invariance principles and the consistency of the variance estimator to show that the bootstrap copies converge jointly with $T_n$.","pith_inferences":["Because the bands are simultaneous over relevant segments, one could invert them to obtain simultaneous confidence intervals for the difference $\\mu_i - \\mu_0$ between fatigue phases, directly testing whether range of motion changes at each point of the stride.","The data example chooses $\\Delta$ from the first and last 5% of the run, which makes the target parameter data-dependent; a rigorous extension would account for that selection in the coverage statement.","For real-time monitoring, one could recompute the bands after each new block of strides; the asymptotic conditions leave open how large $n$ and the block length $L$ must be for nominal coverage in finite samples."],"forward_implications":["The bands in (11) give simultaneous, asymptotically valid coverage over all relevant fatigue segments of one runner, so statements about the range of motion in rest versus fatigued phases hold jointly rather than curve-by-curve.","If no observed jump equals the threshold $\\Delta$, the asymptotic coverage equals $1-\\alpha$; if one jump equals $\\Delta$, coverage stays at least $1-\\alpha-\\beta$, accounting for the preliminary selection of relevant changes.","For the knee-angle data, the bands separate rest, pre-fatigue and fatigue movement patterns, and the fatigue phase shows reduced knee bending at the second peak, matching the protection-mechanism interpretation in the biomechanics literature.","The method turns the biomechanical question into a three-step pipeline---detect changes, keep those with mean jump above $\\Delta$, build bands---so the same recipe applies to hip and ankle angles or to other athletes."],"supporting_citations":[{"why":"Supplies Algorithms 1 and 2 and the consistency theorems for the change point and relevant-set estimators that the proof uses to replace estimated segments by true segments, and provides the original fatigue data analysis.","marker":"Bastian et al. (2024)"},{"why":"Provides Conditions (A1)-(A4) assumed throughout, whose mixing and moment structure underlies the long-run variance consistency and the bootstrap validity.","marker":"Dette and Kokot (2022)"},{"why":"Supplies the weak invariance principle and no-change-point bootstrap result used after estimated segments are replaced by true ones.","marker":"Dette et al. (2020)"},{"why":"Provides the moment inequality for mixing sequences used in the proof of Lemma 1 on consistency of the long-run variance estimator.","marker":"Yoshihara (1978)"}],"fun_headline_variants":["Uniform bands for joint angles across fatigue states","Simultaneous confidence bands for every fatigue phase","One band construction covers all fatigue phases","Bootstrap confidence bands span entire fatigue timeline"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The bands inherit their validity from the unstated conditions (A1)--(A4) of Dette and Kokot (2022) and from fast-enough consistency of the change point and relevant-set estimators; if those assumptions fail, the promised coverage is not guaranteed.","fun_headline_variants_meta":{"raw":{"variants":["Uniform bands for joint angles across fatigue states","Simultaneous confidence bands for every fatigue phase","One band construction covers all fatigue phases","Bootstrap confidence bands span entire fatigue timeline"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000313,"raw_usage":{"total_tokens":1739,"prompt_tokens":867,"completion_tokens":872,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":817}},"tokens_in":483,"tokens_out":872,"duration_ms":7836,"temperature":1.0,"reasoning_tokens":817,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T05:06:29.034046+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate functional time series that satisfy the assumed conditions, with known change points and one jump size exactly equal to $\\Delta$, and compute empirical coverage of the bands over many replications; Theorem 1 predicts coverage at least $1-\\alpha-\\beta$ in that case and exactly $1-\\alpha$ when no jump equals $\\Delta$, so a systematic shortfall would refute the claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies Algorithms 1 and 2 and the consistency theorems for the change point and relevant-set estimators that the proof uses to replace estimated segments by true segments, and provides the original fatigue data analysis."},{"cited_title":"and Kokot, K","cited_arxiv_id":null,"evidence_quote":"Provides Conditions (A1)-(A4) assumed throughout, whose mixing and moment structure underlies the long-run variance consistency and the bootstrap validity."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the weak invariance principle and no-change-point bootstrap result used after estimated segments are replaced by true ones."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the moment inequality for mixing sequences used in the proof of Lemma 1 on consistency of the long-run variance estimator."}],"review_version":1}