{"id":"0c7a0206-b286-467a-8c1b-b996d2c33a44","arxiv_id":"2501.04476","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An L1-norm based change point test for functional time series is introduced, with bootstrap validity, theoretical power comparisons, and a power enhancement against sparse alternatives.","lead":"The paper proposes using the L1 norm, the area between curves, to detect when the average shape of a stream of curves changes, instead of the commonly used L2 or maximum-deviation norms. It argues this choice is more robust to heavy-tailed noise and performs well for spread-out changes, and it adds a safeguard for narrow, spike-like changes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5.1's centering is wrong: under a fixed change at s*, the deterministic part of √n max_s ∥U_n(s)∥_1 is √n s*(1-s*)∥Φ∥_1, so subtracting √n∥Φ∥_1 makes the L1 statistic diverge to -∞; the theoretical power comparison in Corollary 5.2 is therefore not established.","rationale":"The reader's stated weakest assumption focuses on Assumption A4 and beta-mixing, but the reader's rationale already flags the missing s*(1-s*) factor in Theorem 5.1. In my stress-test, this missing factor is the more load-bearing concern for the central claim: Theorem 5.1 and Corollary 5.2 are the paper's theoretical justification for the headline statement that L1 achieves the best performance in a broad range of scenarios, and Theorem 5.1 is false as stated. This is not a matter of overly strong assumptions; it is an internal inconsistency that follows from the paper's own formula for E[U_n(s)]. The proof of Theorem 3.3 also omits details for the H0 case of the bootstrap approximation, which is a genuine gap, but it is plausibly repairable using the sketched argument about the argmax of the limiting Brownian bridge. The false centering, by contrast, is a definite error that must be corrected and the power comparison re-derived. Because the error is localized and the empirical section still provides evidence for L1 in dense-signal settings, a conditional acceptance with required revisions remains appropriate rather than outright rejection.","tokens_in":25959,"tokens_out":10425,"duration_ms":102117,"concrete_test":"Recompute the nonrandom part of T_n^(1) under H1 directly from the paper's formula E[U_n(s)] = (s∧s* - ss*)Φ: evaluate sup_{s∈[0,1]}(s∧s* - ss*) and check whether T_n^(1)-√n∥Φ∥_1 is tight. If the supremum equals s*(1-s*)<1, the centered statistic diverges in probability to -∞, disproving Theorem 5.1 as stated. Then evaluate A1 and A∞ for Φ_c with c=1 and c=10 using the formulas in Theorem 5.1; if both are unchanged, the monotonicity claim in Corollary 5.2 fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5's Theorem 5.1 asserts T_n^(1) - √n∥Φ∥_1 → A1 for fixed Φ, and similarly for the L2 and sup-norm limits. But Section 4's displayed calculation gives E[U_n(s)] = (s∧s* - ss*)Φ, so the non-stochastic maximum of ∥U_n(s)∥_1 is attained at s=s* with value s*(1-s*)∥Φ∥_1. Consequently √n sup_s ∥U_n(s)∥_1 = √n s*(1-s*)∥Φ∥_1 + O_P(1). Since s*(1-s*) < 1 for s*∈(0,1), the statistic T_n^(1)-√n∥Φ∥_1 diverges to -∞ in probability, not to A1. The same missing factor propagates into Eq. (13) and Corollary 5.2, where the centering is q_{1-α,i} - √n∥Φ_c∥_i instead of q_{1-α,i} - √n s*(1-s*)∥Φ_c∥_i, so the theoretical ranking of L1 versus the supremum norm is not supported by the stated result. The adjacent monotonicity claim in Corollary 5.2 is also false for Φ_c(t)=e^{-c(t-0.5)^2}: for every c>0, A1=∫(W(s*)-s*W(1))(t)dt and A∞=(W(s*)-s*W(1))(0.5), so both are independent of c. These are internal inconsistencies, not merely deviations from prior consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops change point detection for the mean of functional time series taking values in L1[0,1]. The authors prove a strong invariance principle for beta-mixing L1-valued sequences (Theorem 2.2), derive the null limit of the L1 cusum statistic (Corollary 3.1), establish consistency of the change-point estimator (Theorem 3.2), and prove asymptotic level and consistency for a block bootstrap test (Theorem 3.3). Section 4 extends the methodology to relevant hypotheses ||mu(1)-mu(2)||_1 <= Delta, with three bootstrap procedures and limit theorems (Theorems 4.1 and 4.2). Section 5 compares L1-, L2-, and sup-norm tests under fixed alternatives: Theorem 5.1 claims T_n^(i) - sqrt(n) ||Phi||_i converges to a nontrivial limit, Eq. (13) translates this into a power formula, and Corollary 5.2 concludes that L1 is best for dense signals while the supremum norm wins for sparse signals; Theorem 5.3 adds a power enhancement for sparse alternatives. Section 6 provides an extensive simulation comparison under light/heavy tails and independence/dependence, and the methodology is applied to Melbourne daily minimum temperature curves.","tokens_in":26335,"tokens_out":36870,"duration_ms":320423,"significance":"If the validity results hold, the paper makes a useful methodological contribution: it provides the first systematic L1-based cusum theory for functional data, including a new strong invariance principle, a bootstrap test with an explicit block construction, relevant-hypotheses tests with level control at Delta = d1, and a power-enhancement mechanism with a user-specified level distortion. The empirical study is thorough (independent and dependent data, light and heavy tails, five signal shapes, two sample sizes, plus a real-data application) and supports the practical recommendation of L1 as a default norm. I find no circularity: the theoretical derivations involve no fitted constants, the bootstrap quantiles are data-dependent by design, and the tuning parameters (block length l_n, distortion level alpha_n, threshold Delta) are standard. However, the advertised theoretical power comparison (Theorem 5.1, Eq. (13), Corollary 5.2) is invalid as stated, for reasons detailed in major comments M1 and M2; the theoretical validation of the 'L1 best' claim needs substantial rework, although the empirical ranking can stand independently.","major_comments":[{"comment":"The centering in Theorem 5.1 contradicts the paper's own Theorem 4.1. Section 4 derives E[U_n(s)] = (s wedge s* - s s*)(mu(1) - mu(2)), and the deterministic factor is maximized at s = s* with value s*(1-s*). Taking Delta = d1 = ||Phi||_1 in the middle case of Theorem 4.1 yields sqrt(n)(sup_s ||U_n(s)||_1 - s*(1-s*) ||Phi||_1) -> T. The stated convergence sqrt(n)(sup_s ||U_n(s)||_1 - ||Phi||_1) -> A1 = T is therefore false whenever s* in (0,1): the left-hand side equals sqrt(n)(s*(1-s*) - 1)||Phi||_1 + O_P(1), which diverges to -infinity in probability. The same missing factor s*(1-s*) affects the L2 and sup-norm statements, whose deterministic peak is again s*(1-s*)||Phi||_i rather than ||Phi||_i. Theorem 5.1 and its proof (which refers back to Theorem 4.1) must be corrected to the centering sqrt(n) s*(1-s*)||Phi||_i before the claims of Section 5 can be evaluated.","section":"Section 5, Theorem 5.1; Section 7.9"},{"comment":"The power comparison is not established even after the centering is corrected. (i) Eq. (13) inherits the error of Theorem 5.1: the shift should be sqrt(n) s*(1-s*)||Phi_c||_i, not sqrt(n)||Phi_c||_i. (ii) With any sqrt(n)-order shift, P ow(i,c) -> 1 as n -> infinity for every fixed c with ||Phi_c||_i > 0; the asserted inequalities P ow(1,0) > P ow(2,0) > P ow(infinity,0) compare three sequences that all converge to 1, so a strict asymptotic ranking requires a large-deviation or local-alternative analysis that the paper does not provide. (iii) The monotonicity claims preceding the corollary are false for the stated family Phi_c(t) = exp(-c(t-0.5)^2): for every finite c, Phi_c > 0 on [0,1], so the set N in Theorem 4.1 is empty and A1 = integral_0^1 (W(s*) - s*W(1))(t) dt is independent of c, while for c > 0 the argmax set is E = {0.5}, so A_infinity = W(s*,0.5) - s*W(1,0.5) is also independent of c; the claims that A1 is increasing in c and A_infinity is decreasing in c are therefore false. Corollary 5.2 should be restated with a nondegenerate comparison (e.g., local alternatives or a fixed-n analysis) or downgraded to an empirical conclusion.","section":"Section 5, Eq. (13), Corollary 5.2"},{"comment":"The bootstrap consistency proof explicitly omits the null-hypothesis case: the text states 'We omit the details as they are not particularly interesting' and only sketches how to find rho with P(||W(s) - sW(1)||_{infinity,1} > a) >= 1 - epsilon and P(max over boundary intervals < a) >= 1 - epsilon. Part 1 of Theorem 3.3, the asymptotic level alpha of the test, is a central validity claim, and the null case is genuinely different from the alternative case because the estimated change point s-hat is not consistent under H0 and the residual correction (mu-hat(2) - mu-hat(1)) is only O_P(n^{-1/2}). The omitted bound for sup_s ||sqrt(n)(S*_n(s) - tilde S*_n(s))||_1 on the event {s-hat in [rho, 1-rho]} must be written out before Theorem 3.3 can be considered proved.","section":"Section 7.3 (proof of Theorem 3.3)"}],"minor_comments":[{"comment":"The text contains garbled passages ('t m ost o ne c hange', 'The hitherto compiled reference all have on thing in common', 'restrospective'); a careful proofread is needed.","section":"Introduction, p. 3"},{"comment":"The indicator in the projection P_N is written as 1{x in [(i-1)/n, i/n)} although the operator is normalized by N on blocks of length 1/N; the second index should be N (blocks [(i-1)/N, i/N)), and as printed the operator is not the intended norm-one projection.","section":"Section 7.2, definition of P_N"},{"comment":"The display 'sqrt(n) ||U_n(s-hat,.)||_infinity -> ||W(s*,.)||_infinity := T' cannot hold as stated: under H0 there is no true s* and the limit would involve the random argmax of the limiting L1-cusum process, while under a fixed alternative the left-hand side diverges. Since Theorem 5.3 only requires eta_n = q^J_{1-alpha_n} -> infinity, the display should be corrected or deleted.","section":"Section 5, power enhancement paragraph before Theorem 5.3"},{"comment":"The integrand 'sgn|U*_n(s,t)|' appears to be a typo for '|U*_n(s,t)|'; the sgn symbol is meaningless as written and the integrand should be the pointwise absolute value.","section":"Section 4, Procedure 3, Eq. (11)"},{"comment":"The displayed formula for Delta-hat_alpha, (d-hat_{infinity,n} - q*_{1-alpha}(n h_n)^{-1/2}) vee 0, does not follow from the definition min{Delta >= 0 | T-hat_{n,Delta} <= q*_{1-alpha}}; solving the inequality gives Delta >= (sup_s ||U_n(s)||_1 - q*_{1-alpha}/sqrt(n)) / (s-hat(1-s-hat)), and the symbols d-hat_{infinity,n} and h_n are not defined at that point.","section":"Remark 4.3"},{"comment":"The quantile ordering q*_{1-alpha,1} < q*_{1-alpha,2} < q*_{1-alpha,infinity} follows from the pointwise norm inequalities ||f||_1 <= ||f||_2 <= ||f||_infinity on [0,1]; attributing it to Jensen's inequality is inaccurate.","section":"Section 5, Eq. (12)"},{"comment":"The proof of Theorem 2.2 is a sketch: the verification of condition (1.19) of Dehling (1983a), in particular the bound E[sup_{|y| <= 2/N} n ||S_n(.) - S_n(.+y)||_1^2] <= N^{-c}, is compressed into a few lines via cotype-2 and Theorem 2.2.4 of van der Vaart and Wellner (1996); since Theorem 2.2 underpins Corollary 3.1 and the proofs of Theorems 3.3 and 4.1, this verification should be written out.","section":"Section 7.2 (proof of Theorem 2.2)"},{"comment":"The sentence 'Similar effects ... can be observed when increasing the factor 0.2 and 0.4 in the definitions of the alternatives (20) to (23)' is unclear; the reader needs to know which parameter (kappa, n, or the signal shape) is varied and how it relates to the numbers reported in Table 3.","section":"Section 6.1, discussion of Table 3"}],"recommendation":"major_revision","confidential_remarks":"The methodological core of Sections 2-4 is plausible and the simulation study is careful, but Section 5 cannot be accepted as is: Theorem 5.1 is internally inconsistent with Theorem 4.1, and Corollary 5.2 rests on false monotonicity claims and a degenerate asymptotic comparison. If the centering is fixed, the honest outcome is that a first-order fixed-alternative comparison cannot rank the three norms since all powers tend to 1 at the same sqrt(n) rate; supporting the theoretical ranking would require local alternatives, a large-deviation analysis, or an explicitly finite-sample statement. If the authors cannot supply such an analysis, the power ranking should be presented as an empirical finding only. The H0 bootstrap proof in Section 7.3 must also be completed. I do not recommend rejection because the L1 methodology itself is a defensible contribution and the empirical evidence for it is strong, but the advertised theoretical comparison needs real rework."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper is worth engaging. The L1 cusum test is new, the strong invariance principle for L1-valued beta-mixing sequences is useful, and the relevant-hypothesis bootstrap procedures plus power-enhancement component are a real contribution. I checked the central test construction; the test statistic and change point estimator are standard machinery, and the claimed convergence of sqrt(n) U_n under H0 to W(s)-sW(1) follows from the invariance principle if that theorem holds. The empirical comparison is also informative, with the expected pattern: L1 robust to heavy tails, sup-norm better for sparse spiky light-tailed signals, L2 in between.\n\nThe soft spots are not cosmetic. Theorem 5.1 as stated is wrong: under a fixed change at s*, E[U_n(s)] = (s^s* - ss*)Phi, so the deterministic maximum of the norm is s*(1-s*)||Phi||_i, not ||Phi||_i. Subtracting sqrt(n)||Phi||_1 therefore makes T_n^(1)-sqrt(n)||Phi||_1 diverge to -infinity, not converge to A1. The same missing factor appears in Eq. (13) and Corollary 5.2, so the theoretical ranking 'L1 best for dense, sup best for sparse' is not supported by the stated results. The adjacent monotonicity claim about A1 increasing and A_infinity decreasing in the sparsity parameter is also false for the Gaussian peaks family: both are independent of c. This is an internal inconsistency in the paper's own setup, not just a disagreement with prior work.\n\nSecond, Theorem 3.3's proof explicitly omits the H0 bootstrap case: the claim 'In the case where H0 holds ... follows along similar lines ... We omit the details' is exactly the case needed for asymptotic level. The H1 case is proven; the level claim is not. That is a genuine gap in the central validity theorem, though likely fillable.\n\nNo code or seeds are provided, so the simulation tables are not directly reproducible. The paper would be stronger if the data-generating code were shipped.\n\nIf I were editing, I would send this to a referee. The L1 proposal is important enough and the invariance principle is formal enough to merit referee time. But acceptance should wait until the Section 5 centering is corrected (or the claims scaled back), the monotonicity claims are fixed, and the H0 bootstrap proof is supplied.","headline":"Genuinely new L1 methodology for functional change point detection, but Section 5's power comparison has a load-bearing centering error and the bootstrap null case is unproved.","tokens_in":26877,"tokens_out":2506,"would_cite":false,"duration_ms":24220,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G10","62G20","62M10","60F17"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims the L1 norm should be the default for functional change point detection, with tests that keep their nominal level and beat L2 and supremum rivals except on sparse, spiky light-tailed signals.","keywords":["functional data","change point detection","L1 norm","cusum statistic","bootstrap","invariance principle","relevant hypotheses","power enhancement"],"falsifier":"Run the L1 block-multiplier bootstrap test under the null on functional data generated as fractional Brownian motion with Hurst exponent close to 1, so the beta-mixing and smoothness assumptions fail, and compare empirical rejection rates to the nominal level; systematic inflation would refute the claimed asymptotic level.","tokens_in":25715,"feed_emoji":"📈","tokens_out":5904,"duration_ms":54892,"temperature":0.7,"pith_summary":"The paper asks which norm should be used when testing for a change in the mean of a functional time series, and argues that the L1 norm is usually the best choice. It constructs a bootstrap test based on the L1 norm of the cusum process, proves that the test has asymptotic level $\\alpha$ and is consistent, and derives the limiting behavior of L1, L2, and supremum norm statistics under a fixed alternative. The L1 methodology also extends to relevant hypotheses that allow small changes of size at most $\\Delta$, for which three bootstrap procedures are proposed and compared. The overall recommendation is L1 as the default norm, with the supremum norm reserved for sparse, spiky signals under light tails and a power-enhancement component to protect L1 against sparse alternatives.","feed_headline":"L1 norm wins for functional change point tests","feed_subtitle":"A bootstrap cusum test in the L1 norm keeps its level, beats L2 and supremum rivals on dense shifts, and handles heavy tails best.","key_machinery":"The central object is the cusum process $U_n(s) = n^{-1}(\\sum_{i=1}^{ns} X_i - s\\sum_{i=1}^n X_i)$ taking values in $C([0,1],L^1)$, with the test statistic $\\hat T_n = \\sqrt{n}\\sup_{s\\in[0,1]}\\|U_n(s)\\|_1$. The argument is carried by a new strong invariance principle for $L^1$-valued $\\beta$-mixing sequences, which uses the cotype-2 property of $L^1$ and projection approximations to embed the partial sums next to an $L^1$-valued Brownian motion. Directional Hadamard differentiability of the supremum-of-$L^1$-norm functional then yields the asymptotic distributions under both null and alternative, and a block multiplier bootstrap provides the quantiles.","core_discovery":"The paper establishes that the L1-norm cusum bootstrap test has asymptotic level alpha under the null and power tending to one under any fixed alternative, and that under a fixed alternative the L1 statistic grows as sqrt(n) times the L1 norm of the mean difference. For dense signals this gives L1 a theoretical advantage over both L2 and supremum norms, while the supremum norm is shown to be better for sufficiently sparse and spiky alternatives. The paper also introduces a power-enhancement term that detects spatially localized changes the plain L1 test misses, at a user-specified cost in size distortion, and validates the whole methodology through simulations and a temperature-curve data example.","pith_inferences":["A testable extension is an online monitoring version: the paper only treats retrospective at-most-one-change detection, but the same cusum and bootstrap ingredients are natural candidates for sequential monitoring.","Because the $L^1$ statistic is a sum of absolute integrated errors, the power comparison may extend to functional data with infinite second moments, where the paper's moment assumptions fail.","The user-specified tolerated distortion $\\alpha_n$ in the power enhancement suggests a data-driven choice of $\\alpha_n$ based on estimated signal sparsity, which the paper leaves implicit.","The paper's analysis of the Gaussian class $\\Phi_c(t)=e^{-c(t-0.5)^2}$ predicts a crossover between $L^1$ and supremum power; systematic simulation across covariance operators $C$ could map where that crossover sits."],"forward_implications":["The $L^1$ bootstrap test can serve as a default for retrospective functional change point detection: it keeps the nominal level and detects dense mean shifts without dimension reduction.","Under a fixed alternative, the $L^1$ statistic separates from its null quantiles at rate $\\sqrt{n}$ according to the signal's $L^1$ norm, so power improves as the signal spreads over the domain.","The power-enhancement component detects spatially concentrated mean shifts that the plain $L^1$ test misses, inflating the asymptotic level only by the user-chosen amount $\\alpha_n$.","For relevant hypotheses with threshold $\\Delta$, the bootstrap procedure based on the upper-bound statistic (11) keeps the nominal level in moderate samples and gives an interpretable area-between-curves threshold.","The strong invariance principle for $L^1$-valued $\\beta$-mixing sequences is itself a reusable tool for other inference problems on integrable functional data."],"supporting_citations":[{"why":"Supplies the supremum-norm methodology and the relevant-hypotheses framework that the paper compares against.","marker":"Dette et al. (2020)"},{"why":"Supplies the L2 sequential block bootstrap test used as the L2 comparison.","marker":"Sharipov et al. (2016)"},{"why":"Provides the cotype and Gaussian concentration results that underpin the L1 invariance principle.","marker":"Ledoux and Talagrand (1991)"},{"why":"Provides the strong invariance principle for weakly dependent Banach-valued variables that Theorem 2.2 adapts to L1.","marker":"Dehling (1983a)"},{"why":"Provides the tightness criterion used to control moduli of continuity in the bootstrap proofs.","marker":"van der Vaart and Wellner (1996)"},{"why":"Provides the conditional weak convergence lemma linking bootstrap consistency to joint weak convergence.","marker":"Bücher and Kojadinovic (2019)"},{"why":"Supplies the notion of a power enhancement component that the paper adapts for sparse alternatives.","marker":"Fan et al. (2015)"}],"fun_headline_variants":["L1 cusum beats L2 and sup for functional shifts","Power-boosted L1 test top for functional change points","L1 norm outperforms L2 and sup for functional change","L1 norm best for functional change point power"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole construction assumes the noise sequence forgets its past quickly and has a Gaussian counterpart whose L1 distance between shifted versions decays at a polynomial rate; if the data have rough or long-memory noise, neither the limiting null distribution nor the bootstrap's validity is guaranteed.","fun_headline_variants_meta":{"raw":{"variants":["L1 cusum beats L2 and sup for functional shifts","Power-boosted L1 test top for functional change points","L1 norm outperforms L2 and sup for functional change","L1 norm best for functional change point power"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001607,"raw_usage":{"total_tokens":6326,"prompt_tokens":800,"completion_tokens":5526,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":416,"completion_tokens_details":{"reasoning_tokens":5469}},"tokens_in":416,"tokens_out":5526,"duration_ms":41207,"temperature":1.0,"reasoning_tokens":5469,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:32:55.944459+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the L1 block-multiplier bootstrap test under the null on functional data generated as fractional Brownian motion with Hurst exponent close to 1, so the beta-mixing and smoothness assumptions fail, and compare empirical rejection rates to the nominal level; systematic inflation would refute the claimed asymptotic level.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the notion of a power enhancement component that the paper adapts for sparse alternatives."}],"review_version":1}