{"id":"abc031c8-3af2-4e78-98d5-2aa81b794b57","arxiv_id":"2412.20858","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A unified B-spline CUMSUM framework for detecting and dating mean-function structural breaks in functional time series, with theory and inference valid from sparse to dense sampling.","lead":"This paper develops statistical tests that detect abrupt changes in curves observed over time, whether each curve has only a few measurements or is measured densely. It also locates the break date and builds confidence bands for the size of the jump, with applications to electricity prices and temperatures.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.1's unified guarantee is conditional on (A6), but the BIC knot selector in §5.1 is not shown to satisfy those rate conditions, so the implemented test may not have the advertised size.","rationale":"The reader's weakest_assumption is exactly (A6), and the same condition is the most load-bearing element in the paper's central claim. Theorem 3.1 asserts a Gaussian approximation with no restriction on the relationship between n and N_i, but that claim is qualified by the coupled rate conditions in (A6), which constrain J_n, r, and E(N^{-1}). The paper's own implementation, however, selects J_n via a BIC criterion whose admissible rates are not derived from (A6). This creates a genuine gap between the proven oracle result and the procedure used in the simulations and applications. This is not a disagreement with the consensus or a mere technicality: if the BIC-selected J_n falls outside the (A6) window, the test's size is not guaranteed, and the phase-transition boundaries in Corollary 3.1 may not hold for the implemented statistic. I therefore agree with the reader's identification of (A6) as the weakest assumption, and the concrete test above would determine whether this concern actually affects the reported finite-sample performance. The reader's CONDITIONAL verdict already reflects the need for additional verification; my analysis does not move that verdict, so UNCHANGED is appropriate.","tokens_in":175,"tokens_out":14180,"duration_ms":156643,"concrete_test":"Run a simulation study for Setting (1) (sparse, N_i ~ U{3,...,6}, n=200 and 400, 500 Monte Carlo replications) and record the BIC-selected J_n for every replication. Compute the empirical E(N^{-1}) and the implied r for the standardized Laplace errors (e.g., r=4,5,6). For each r, check whether the observed J_n falls in the (A6)-admissible range implied by the two branches, including verification that the moment-bias product tends to 0 for the sequence. Then fix the same data-generating process and rerun the size/power analysis with J_n forced to a value inside the (A6) window (e.g., J_n = floor(n^{1/(2q*+1)}) with q*=4) and to a value clearly outside (e.g., J_n = 2 * that value). If the empirical size is close to 0.05 inside but inflates or undercoverage occurs outside, the BIC implementation is not covered by Theorem 3.1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Assumption (A6) is the only place coupling the knot number J_n, the moment order r, and the sampling intensity E(N^{-1}); Theorem 3.1's 'no restriction on n vs N_i' is valid only for J_n inside that coupled rate window. The implementation in Section 5.1 selects J_n by minimizing BIC over a fixed range (e.g., [0.5(n\\bar N)^{1/9}, (n\\bar N)^{1/7}] for n=200, sparse Setting (1)). No lemma shows the BIC-optimal J_n satisfies either branch of (A6), nor that the moment-bias product in (A6) tends to 0. In Setting (1), where E(N^{-1}) is bounded away from 0, the BIC range for n=200 gives J_n roughly 1–2; whether this lies in the admissible region depends on the effective smoothness q* and an available moment order r. If r is only 3, the displayed moment-bias expression grows for typical J_n, so the Gaussian approximation in Theorem 3.1 can fail exactly in the sparse regime the paper claims to unify. Thus the central result is proven only for an oracle statistic with a user-specified J_n satisfying (A6), not for the automatic BIC procedure whose size and power are reported in Tables 1–3.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a unified methodology for testing and estimating a structural break in the mean function of functional time series whose trajectories are observed at irregular, possibly sparse locations. The authors construct a smoothed CUMSUM process via B-spline estimation of partial and global mean functions, then define L-infinity and L-2 test statistics. The main theoretical results are a Gaussian approximation for the null distribution (Theorem 3.1), local alternative power analysis (Theorem 3.2), convergence rates for break-point estimators (Theorem 3.3), and a simultaneous confidence band for the jump magnitude (Theorem 3.4). Corollary 3.1 gives phase-transition boundaries in terms of the sampling intensity E(N^{-1}). The paper also contains an extensive simulation study across four sampling schemes and three jump shapes, plus applications to German electricity price data and Sydney temperature data.","tokens_in":22922,"tokens_out":5444,"duration_ms":49938,"significance":"If the theoretical results are correct, the paper would provide a genuinely unified treatment of structural-break testing from sparse to dense functional data, and the L-infinity statistic is a useful complement to the existing L-2-based procedures. The phase-transition analysis in Corollary 3.1 is a substantive contribution, and the simulation study is unusually thorough, covering several sampling regimes, error distributions, and jump shapes. The main novelty—Gaussian approximation of a smoothed CUMSUM process without a restriction on the relationship between n and Ni—is attractive and would be of broad interest to the functional data and change-point communities. However, the posted arXiv version does not include the supplementary proofs, and the gap between the theoretical Assumption (A6) and the implemented BIC knot-selection procedure means that the paper as written does not yet establish the advertised guarantees for the practical algorithm.","major_comments":[{"comment":"The null-distribution result is stated for a user-specified number of knots J_n satisfying the coupled rate conditions in (A6), but the implementation selects J_n by minimizing BIC over the range [min{0.5(n\\bar N)^{1/9}, 0.5n^{1/8}}, max{(n\\bar N)^{1/7}, n^{1/6}}]. No lemma or argument shows that the BIC-selected J_n satisfies either branch of (A6), nor that the moment-bias product in (A6) tends to zero. In sparse Setting (1), E(N^{-1}) is bounded away from zero and, for n=200, the BIC range yields J_n of order 1--2; whether that lies in the admissible region depends on the effective smoothness q* and the available moment order r. If r is small (for instance r=3), the displayed moment-bias expression can grow, so the Gaussian approximation in Theorem 3.1 can fail exactly in the sparse regime the paper claims to unify. Consequently, the empirical sizes and powers in Tables 1--3 are reported for a procedure whose asymptotic guarantees are not established; the paper proves the result for an oracle statistic with J_n in (A6), not for the automatic BIC procedure actually used.","section":"Section 5.1, Assumption (A6), Theorem 3.1"},{"comment":"The text says that 'Since no restrictions are imposed on the relationship between n and Ni in Assumptions (A1)-(A6), Theorem 3.1 contains scenarios with arbitrary sampling schemes from sparse to dense.' This is inaccurate: Assumption (A6) restricts J_n, n, E(N^{-1}), and r jointly, for example by requiring J_n^{-q*-1/2} n^{1/2} E^{-1/2}(N^{-1}) log^{1/2} n = O(1) when E(N^{-1}) J_n \\gg 1. Thus 'arbitrary sampling schemes' is too strong; the theorem holds only for sampling intensities and knot numbers lying in the coupled window described by (A6). The abstract's claim of being 'adaptable to arbitrary sampling schemes' should be weakened, or the authors should show that (A6) is automatically satisfied under the stated model assumptions, which is not done.","section":"Theorem 3.1 and the paragraph following it"},{"comment":"The technical proofs of Theorems 3.1--3.4 are deferred to a supplementary file that is not included in the posted arXiv v1. Because the paper's central claims rest entirely on these Gaussian approximation results—and the reader can neither check the derivations, the precise conditions under which the phase-transition boundaries in Corollary 3.1 hold, nor the treatment of the estimated covariance matrix \\hat\\Sigma—the posted version is not self-contained. For a revised submission, the supplement should be made available, or the main text should summarize the key intermediate lemmas and indicate where each part is proven.","section":"Supplement (end of Section 1 and beginning of Section 3)"}],"minor_comments":[{"comment":"The abstract contains the duplicated phrase 'test statistics statistics'; also in Section 1, 'the discrete girds' should read 'the discrete grids.'","section":"Abstract and Section 1"},{"comment":"In the local alternative display, the expression 'max{n√ n−1, p n−1E(N −1)Jn}' is confusing because √ n^{-1} equals n^{-1/2}; writing n^{-1/2} explicitly would improve readability.","section":"Theorem 3.2"},{"comment":"The final sentence says 'bkn,∞ becomes more accurate than bkn,∞ with finite sample sizes,' which should presumably read 'bkn,∞ becomes more accurate than bkn,2.'","section":"Section 3.2"},{"comment":"The caption refers to 'SCRs' but the text and method name use 'SCB'; please correct the abbreviation.","section":"Table 3 caption"},{"comment":"Several references are incomplete or informal, for instance 'Bai, L., Hu, Q., and Wu, W. (2024+)' and 'Dette, H. and Wu, W. (2024+)' are listed as 'Manuscript' without an arXiv or DOI identifier; please update these entries.","section":"References"},{"comment":"The column header 'Linfinity' should be typeset as 'L∞' for consistency with the rest of the paper.","section":"Table 2 header"}],"recommendation":"major_revision","confidential_remarks":"The manuscript makes an ambitious and potentially useful claim, and the simulation study is extensive. The two main obstacles to acceptance are the gap between Assumption (A6) and the BIC knot-selection procedure, and the absence of the supplementary file with the proofs in the posted arXiv version. The phase-transition corollary is a nice contribution and could become the basis for a strong paper if the connection to the implemented procedure is established. The authors should also carefully document how Theorems 3.1 and related results depend on the machinery in their companion papers (Cai and Hu 2024a,b), since the current text appears to inherit substantial Gaussian-approximation tools from those works without making the exact dependency explicit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short take: this is a genuinely useful unification of break-point testing for functional time series under sparse-to-dense sampling. One B-spline smoothed CUMSUM framework, both L∞ and L2 tests, a phase-transition corollary, break-point estimators, and a simultaneous band for the jump. If the supplement is right, it goes beyond Aue et al. (2018) and Bai et al. (2024) to cover irregular, unequally sampled designs. That is a real contribution.\n\nWhat the paper does well: the theoretical architecture is sensible. It uses Mies and Steland's sequential Gaussian approximation and Chernozhukov-type anti-concentration, and the derived limits reduce to known results in fully observed and semi-dense special cases—that is a good sign. The simulations are extensive: four sampling schemes, three jump shapes, three error distributions, n=200 and 400. The empirical size is near nominal, power patterns make sense, and the L∞/L2 trade-off for spike-like versus smooth jumps is a nice practical insight.\n\nSoft spots, in order of importance. First, proofs are in the supplemental material, which is not part of the posted arXiv v1. So the central derivations cannot be checked from what's here. Second, the stress-test note is on target: the unified guarantee in Theorem 3.1 is conditional on (A6), a coupled knot-moment-sampling condition, and the BIC knot selector in Section 5.1 is not shown to satisfy it. The theorem as written covers an oracle J_n in an admissible window; the tables report an automatic BIC choice. That's a genuine theory-practice gap. It's probably patchable—for the smooth mean functions and normal errors in the simulations, one can take the moment order r large enough—but as written, the size guarantees in Tables 1–2 don't directly follow from the theorem. Third, the abstract says 'arbitrary sampling schemes,' but the main assumptions require random i.i.d. design (X_{ij} with a density, N_i i.i.d.). Both applications use fixed designs (daily electricity demand, day-of-year temperature), which fall outside the theorem. The simplified fixed-design estimator in Section 4.1 has no supporting theorem. This is an overreach, not a fatal flaw.\n\nBottom line: solid, significant methodology, honestly presented, with a couple of gaps between the theory and the implementation. I'd send it to a serious referee and expect they'll want to see the supplement and a fix or explicit discussion of the BIC condition. I'd also bring it to a reading group.","headline":"Unified break-point test for functional time series from sparse to dense; real contribution, but the automatic knot choice runs ahead of the theory and the fixed-design applications are outside the theorem.","tokens_in":23463,"tokens_out":5358,"would_cite":true,"duration_ms":51521,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G10","62G20","62G08","62M10","62R10"],"pacs":[],"model":"deepseek-v4-flash","headline":"A B-spline-smoothed CUMSUM process yields functional break-point tests whose null distributions hold uniformly across sparse, semi-dense, and dense sampling, alongside a phase-transition boundary and post-break inference.","keywords":["functional time series","structural breaks","change point detection","B-spline smoothing","CUMSUM process","sparse to dense functional data","phase transition","Gaussian approximation"],"falsifier":"Simulate the null model (1.1) in a sparse regime where (A6) is violated—for example $E(N^{-1})J_n \\gg 1$ while $J_n^{-q^*-1/2} n^{1/2}E^{-1/2}(N^{-1})a_n$ diverges—and compare the empirical distribution of $T_n$ with the quantile of $\\sup_{\\epsilon\\le t\\le1-\\epsilon}\\sup_{x\\in[0,1]}|Z_n(t,x)|/\\sqrt{B^{\\top}(x)\\Sigma B(x)}$; a persistent size distortion would show these rate conditions are doing real work.","tokens_in":22397,"feed_emoji":"📊","tokens_out":10829,"duration_ms":99448,"temperature":0.7,"pith_summary":"This paper tries to establish that the same change-point tests for the mean of a functional time series work no matter how thinly or densely each curve is sampled. Previous tests either require fully observed trajectories, regular grids, or sampling frequencies that grow at a prescribed rate relative to the number of curves. The authors construct L∞ and L2 statistics from a B-spline-smoothed CUMSUM process and prove, under a common set of assumptions, that both have asymptotically correct size across all sampling regimes and a Gaussian-field null limit. They also derive the boundary where the effective limit switches from a noise-dominated sparse regime to a dependence-dominated dense regime, give convergence rates for the estimated break date, and build simultaneous confidence bands for the jump function. If true, one procedure could replace a patchwork of regime-specific tests.","feed_headline":"Break-point tests now span sparse to dense curve data","feed_subtitle":"B-spline smoothing yields L2 and L∞ tests with correct size in every sampling regime, and pinpoints the sparse-dense boundary.","key_machinery":"The load-bearing object is the standardized smoothed CUMSUM process $Q_n(t,x)=([nt]\\hat m(t,x)/\\sqrt n - \\sqrt n\\,t\\,\\hat m(x))/\\sqrt{B^{\\top}(x)\\Sigma B(x)}$, where $\\hat m(t,\\cdot)$ and $\\hat m(\\cdot)$ are B-spline least-squares estimates of the partial and global mean with per-curve weight $N_i^{-1}$. Under the null this process is approximated by a zero-mean Gaussian field $Z_n(t,x)$ with covariance $(\\min\\{t,t'\\}-tt')B^{\\top}(x)\\Sigma B(x')$, and the normalizer $\\sqrt{B^{\\top}\\Sigma B}$ keeps the statistic from diverging when few points per curve make the noise term dominate. The proof machinery combines sequential Gaussian approximation for dependent vector-valued series, anti-concentration bounds for maxima of Gaussian vectors, and spline basis properties; the decomposition $\\Sigma=\\Sigma_1+\\Sigma_2$ is what turns the sparse-to-dense phase transition into a comparison of orders.","core_discovery":"On the paper's own terms, the central discovery is that, under assumptions (A1)-(A6), $\\sup_z |P(T_n \\le z) - P(\\sup_{\\epsilon\\le t\\le1-\\epsilon}\\sup_{x\\in[0,1]} |Z_n(t,x)|/\\sqrt{B^{\\top}(x)\\Sigma B(x)} \\le z)| \\to 0$ with no restriction on how $n$ and the per-curve counts $N_i$ are related; the same holds for $S_n$ with the integral of the squared field under the slightly lighter (A6'). Corollary 3.1 then splits the regimes by comparing $E(N^{-1})$ with $n^{1/(2q^*)}$: sparse sampling is dominated by the noise term $\\Sigma_2$, dense by the dependence term $\\Sigma_1$, and the overlap is a mixed semi-dense regime. Under local alternatives the minimal detectable jump is of order $\\max\\{\\sqrt{\\log n/n}, \\sqrt{E(N^{-1})J_n\\log n/n}\\}$ in sup-norm and its $L^2$ analogue, so the tests are consistent whenever the jump exceeds the regime-dependent threshold. The same construction yields break-point estimators and an asymptotically valid simultaneous confidence band for the jump magnitude $\\Delta(\\cdot)$.","pith_inferences":["Editorial inference — The BIC-based knot choice used in the simulations and applications is not proved to satisfy (A6), so the guarantees proven here may not cover the exact procedure as implemented.","Editorial inference — The threshold $E^{-1}(N^{-1}) \\asymp n^{1/(2q^*)}$ coincides with the known estimation phase transition for mean functions, suggesting a general rule: change-point detection pays the nonparametric smoothing price when curves are sparse and recovers the parametric $n^{-1/2}$ rate once curves are dense.","Editorial inference — The same normalized smoothed CUMSUM construction could be adapted to test breaks in covariance operators or locally stationary functional series; the authors list both as future directions rather than established results.","Editorial inference — Taken together, the simulations imply a practical rule the paper does not formally state: use the L∞ test for spiky jump functions and the L2 test for flat ones, and interpret a rejection by either as evidence of a break."],"forward_implications":["The same two test statistics can be applied without first classifying the sampling scheme; the normalizer $\\sqrt{B^{\\top}(x)\\Sigma B(x)}$ automatically tunes the limit from noise-dominated to dependence-dominated.","The phase-transition boundary separates sparse ($E^{-1}(N^{-1}) \\ll n^{1/(2q^*)}$) from dense ($E^{-1}(N^{-1}) \\gg n^{1/(2q^*)}$) sampling, up to logarithmic factors, with a mixed semi-dense regime in between.","Under local alternatives, the tests have nontrivial power once the jump exceeds $n^{-1/2}$ in dense or semi-dense settings and $\\sqrt{E(N^{-1})J_n/n}$ in sparse settings, and power tends to 1 when the jump is larger.","The $L^2$-based break-point estimator is within $O_p(\\max\\{1, J_n E(N^{-1})\\})$ of the true break index, which is optimal in the dense case; the $L^\\infty$ estimator has a slower proven rate but better finite-sample accuracy for sharp jumps.","The two-step simultaneous confidence band for the jump magnitude has asymptotically correct coverage across all sampling regimes, so post-break inference does not require a separate dense-data method."],"supporting_citations":[{"why":"It supplies the fully observed-trajectory L2 CUMSUM benchmark whose limiting distribution the new statistics match in the dense limit.","marker":"Aue et al. (2018)"},{"why":"It provides the earlier semi-dense fixed-design detector without pre-smoothing that the paper uses as a baseline for its sparse-to-dense claim.","marker":"Bai et al. (2024)"},{"why":"It gives the sequential Gaussian approximation for dependent vector-valued series that handles the partial-sum spline coefficients.","marker":"Mies and Steland (2023)"},{"why":"It supplies the comparison and anti-concentration bounds for maxima of Gaussian vectors that control the sup-norm approximation error.","marker":"Chernozhukov et al. (2015)"},{"why":"It establishes the sparse/semi-dense/dense taxonomy and phase transitions in functional mean estimation that the paper translates into change-point boundaries.","marker":"Zhang and Wang (2016)"},{"why":"It provides the earlier B-spline phase-transition analysis for simultaneous mean inference whose knot-and-regime translation strategy is adapted here.","marker":"Cai and Hu (2024a)"},{"why":"It is the existing break-point estimation method for sparse and dense functional data that the paper improves and uses as a rate comparison.","marker":"Madrid Padilla et al. (2022)"},{"why":"It gives the optimal mean-estimation rates under weighted smoothing that motivate the B-spline estimator and its per-curve weights.","marker":"Cai and Yuan (2011)"}],"fun_headline_variants":["Break-point tests adapt to sparse and dense curves","New tests find jumps in sparse or dense functional data","Sparse-to-dense break detection with B-spline CUMSUM","Phase transition revealed for break-point detection in curves","B-spline tests for abrupt shifts in functional time series"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theory holds only if the spline order, moment conditions, and per-curve sampling intensities satisfy a list of coupled rate restrictions, and the data-driven BIC knot selection used in the applications is not shown to meet them.","fun_headline_variants_meta":{"raw":{"variants":["Break-point tests adapt to sparse and dense curves","New tests find jumps in sparse or dense functional data","Sparse-to-dense break detection with B-spline CUMSUM","Phase transition revealed for break-point detection in curves","B-spline tests for abrupt shifts in functional time series"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000162,"raw_usage":{"total_tokens":1244,"prompt_tokens":958,"completion_tokens":286,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":207}},"tokens_in":574,"tokens_out":286,"duration_ms":3072,"temperature":1.0,"reasoning_tokens":207,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:08:33.034271+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the null model (1.1) in a sparse regime where (A6) is violated—for example $E(N^{-1})J_n \\gg 1$ while $J_n^{-q^*-1/2} n^{1/2}E^{-1/2}(N^{-1})a_n$ diverges—and compare the empirical distribution of $T_n$ with the quantile of $\\sup_{\\epsilon\\le t\\le1-\\epsilon}\\sup_{x\\in[0,1]}|Z_n(t,x)|/\\sqrt{B^{\\top}(x)\\Sigma B(x)}$; a persistent size distortion would show these rate conditions are doing real work.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the fully observed-trajectory L2 CUMSUM benchmark whose limiting distribution the new statistics match in the dense limit."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It provides the earlier semi-dense fixed-design detector without pre-smoothing that the paper uses as a baseline for its sparse-to-dense claim."},{"cited_title":"and Steland, A","cited_arxiv_id":null,"evidence_quote":"It gives the sequential Gaussian approximation for dependent vector-valued series that handles the partial-sum spline coefficients."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the comparison and anti-concentration bounds for maxima of Gaussian vectors that control the sup-norm approximation error."},{"cited_title":"and Wang, J.-l","cited_arxiv_id":null,"evidence_quote":"It establishes the sparse/semi-dense/dense taxonomy and phase transitions in functional mean estimation that the paper translates into change-point boundaries."},{"cited_title":"M., Wang, D., Zhao, Z., and Yu, Y","cited_arxiv_id":null,"evidence_quote":"It is the existing break-point estimation method for sparse and dense functional data that the paper improves and uses as a rate comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It gives the optimal mean-estimation rates under weighted smoothing that motivate the B-spline estimator and its per-curve weights."}],"review_version":1}