{"id":"ba2c7121-3566-4860-a4c5-65a69ce05277","arxiv_id":"2501.00118","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A bootstrap portmanteau test for white noise in multivariate locally stationary functional time series, backed by a Gaussian approximation for maxima of degenerate U-statistics with an increasing number of lags.","lead":"Statisticians propose a test for whether a multivariate time series of curves is pure noise even when the data-generating process drifts over time. It aggregates autocovariance evidence over many lags without compressing the curves, and a bootstrap delivers critical values.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.2's rate condition cannot hold under the paper's own illustrative tuning: with τ≍n^{-2/5} and s_n=o(n^{1/55}), s_n/√(nτ^3)=s_n n^{1/10}→∞ rather than →0.","rationale":"I read the paper in good faith. The Gaussian approximation argument is detailed and the geometric physical-dependence assumption in Definition 2.1 and Assumption 3.1(2) is strong but explicit; Examples 2.1 and 2.2 are designed to satisfy it, and the cumulant bounds in Remark 3.1 and Lemma B.2 are internally consistent under that assumption. So the reader's weakest assumption, geometric decay, is a genuine scope limitation rather than an unexplained internal gap. The more concrete and load-bearing issue is the rate inconsistency in the consistency theorems: Theorem 3.2's hypothesis s_n/√(nτ^3)→0 cannot hold for the rates that Theorem 3.3 explicitly advertises. Since Theorem 3.4 requires the conditions of Theorems 3.1–3.3 jointly, the paper currently lacks a demonstrated admissible tuning regime for its consistency claim. The proof of Theorem 3.2 does not seem to use the suspicious condition, so the most plausible resolution is a typo; nevertheless the v1 text needs correction. This supports keeping the CONDITIONAL verdict rather than rejecting the paper: the core asymptotic construction appears salvageable, but the stated conditions need repair and the claimed rate illustration is wrong. The absence of the advertised real-data analysis and pinned code are additional manuscript defects, but they are not the central mathematical claim.","tokens_in":68795,"tokens_out":17718,"duration_ms":201089,"concrete_test":"Independently re-derive Theorem 3.2's proof from Appendix C.1 without invoking the condition s_n/√(nτ^3)→0. If the conclusion Q_n/√(s_n)(nτ)^d→∞ still follows from (A.1)–(A.4), then the condition is a typo; in that case, locate a corrected condition and verify a concrete rate (e.g., τ≍n^{-1/3+ε} with small ε>0, s_n=logn) that simultaneously satisfies Theorems 3.1, 3.2, and 3.3. If the proof genuinely requires s_n/√(nτ^3)→0, exhibit any sequence (τ_n, s_n, L_n, b_n, N_n) satisfying all hypotheses of Theorems 3.1–3.3; otherwise the consistency claim of Theorem 3.4 is unproven.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the bootstrap test's asymptotic validity (Theorems 3.1–3.4). The proof machinery is largely coherent under the stated geometric dependence conditions, but Theorem 3.2 contains a rate condition that is incompatible with the rates advertised for Theorem 3.3. Specifically, Theorem 3.2 requires s_n/√(nτ^3)→0. The text immediately after Theorem 3.3 says its conditions are satisfied for N=O(n^λ), b≍n^{-1/4}, τ≍n^{-2/5}, s_n=o(n^{1/55}), L≍n^{1/5}. For these rates, nτ^3=n^{-1/5}, so √(nτ^3)=n^{-1/10}, and s_n/√(nτ^3)=s_n n^{1/10}, which diverges (indeed o(n^{13/110})). Thus the only explicit regime claimed to satisfy Theorem 3.3 violates a hypothesis of Theorem 3.2. Theorem 3.4 invokes both theorems, so no admissible rate satisfying all hypotheses is exhibited. The proof of Theorem 3.2 in Appendix C.1 does not appear to use the condition s_n/√(nτ^3)→0; it relies instead on bounds like s_nτ=o(1). This suggests the condition may be a typo, but as written the consistency theorem has a gap: either the condition is unnecessary and should be removed/corrected, or a genuinely compatible rate must be supplied. This is more immediately load-bearing than the geometric-dependence assumption, which is an explicit scope restriction and is satisfied by the paper's examples.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a fully functional portmanteau-type test for the null hypothesis that a multivariate locally stationary functional time series is white noise, without dimension reduction and with a diverging number of lags. The test statistic Q_n aggregates nonparametric estimates of squared Frobenius norms of autocovariance matrices over a time–argument grid, and critical values are obtained from a difference-based block multiplier bootstrap. The main results are Theorem 3.1 (Gaussian approximation for a maximum of degenerate U-statistics with increasing dimension and dependence), Theorem 3.2 (divergence of the statistic under an alternative), Theorem 3.3 (bootstrap consistency under null and alternative), and Theorem 3.4 (asymptotic level and consistency of the bootstrap test). The supplement contains detailed proofs, cumulant bounds, and rate calculations.","tokens_in":69157,"tokens_out":20216,"duration_ms":184927,"significance":"If the results are valid, the paper fills a genuine gap: existing white-noise tests for functional time series are mostly univariate, stationary, or based on dimension reduction, whereas this paper treats multivariate locally stationary functional data without a Karhunen–Loève projection and allows the number of lags to grow. The Gaussian approximation for maxima of degenerate U-statistics of second-order functional time series with increasing dependence appears to be new and of independent interest. The paper is careful in stating explicit rate conditions, and the appendix provides a substantial and structured proof apparatus. The main reservation is that a rate condition in Theorem 3.2 is not satisfiable as stated, so the consistency part of the paper currently lacks an admissible rate regime.","major_comments":[{"comment":"Theorem 3.2 requires s_n / sqrt(n tau^3) -> 0. Since tau -> 0 and n tau^3 -> 0 under any admissible bandwidth, sqrt(n tau^3) -> 0; with s_n -> infinity the ratio cannot tend to 0 for any admissible sequence. The paper's own rate proposal after Theorem 3.3 (N=O(n^lambda), b ~ n^{-1/4}, tau ~ n^{-2/5}, s_n = o(n^{1/55}), L ~ n^{1/5}) gives sqrt(n tau^3) = n^{-1/10}, so s_n / sqrt(n tau^3) = s_n n^{1/10} -> infinity, violating the condition. Since Theorem 3.4 invokes Theorem 3.2, no rate regime satisfying all hypotheses is exhibited. The proof of Theorem 3.2 in Appendix C.1 does not use this condition; it relies instead on bounds such as s_n tau^{-2/s} = o(sqrt(n tau)), suggesting the condition may be a typographical error. This must be corrected and a compatible rate regime supplied before the consistency claims are supported.","section":"Section 3, Theorem 3.2 and the paragraph after Theorem 3.3"}],"minor_comments":[{"comment":"The arXiv abstract promises a real data analysis of energy consumptions and points to an R package, but the v1 manuscript contains only simulation studies and no real-data section or package reference; the abstract should be aligned with the actual content.","section":"Abstract and Section 4"},{"comment":"The initial statistic in (2.6) is written without an absolute value and without the normalization that appears in Q_n in (2.17); the two displays should define the same statistic consistently.","section":"Section 2.2, equations (2.6) and (2.17)"},{"comment":"The alternative hypothesis is stated in terms of |tr(Gamma_k(t,u))|^2, while the test statistic aggregates estimates of |Gamma_k(t,u)|_F^2. Because |Gamma_k|_F^2 >= p^{-1}(tr Gamma_k)^2, the stated alternative is sufficient for the proof, but the connection should be made explicit and the hypothesis should be formulated in terms of the quantity actually estimated.","section":"Section 3, equation (3.3)"},{"comment":"The statement 'Qn/√sn(nτ)d → ∞' should be parenthesized as Q_n/(sqrt{s_n} (n tau)^d) -> infinity; in addition, in the proof around (A.3) the squared Frobenius norm appears without the opening |·| symbol.","section":"Section 3, Theorem 3.2 and Appendix C.1"},{"comment":"The phrase 'there exists a sequence of Gaussian process (Z_k)... ∈ R^{N(n−2⌈nτ⌉+1)}' should specify that each Z_k is an N(n−2⌈nτ⌉+1)-dimensional vector, not that the whole sequence lies in that space.","section":"Section 3, Theorem 3.1"},{"comment":"Example 2.2 says condition (2) of Assumption 3.1 holds 'with order 2', but Assumption 3.1(2) requires s* >= 4; this discrepancy should be reconciled.","section":"Example 2.2 and Assumption 3.1(2)"}],"recommendation":"major_revision","confidential_remarks":"The rate-condition error in Theorem 3.2 appears fixable, since the proof in Appendix C.1 indicates the stated condition is stronger than what is actually used. I therefore recommend major revision rather than rejection. The missing real-data section should also be resolved, either by adding it or by removing it from the abstract."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core result is solid and the paper is worth engaging with. It extends Bücher–Dette–Heinrichs from univariate stationary/locally stationary functional series to multivariate locally stationary series, removes dimension reduction, allows the number of lags to diverge, and proves a new high-dimensional Gaussian approximation for degenerate U-statistics under moment conditions rather than sub-Gaussian tails. The difference-based block multiplier bootstrap is a natural fit for the increasing dependence, and the appendix is long and detailed; the cumulant bounds in Remark 3.1 and Lemma B.2 are the technical heart and look correct.\n\nThe one real problem is the rate condition in Theorem 3.2. It requires s_n / sqrt(n tau^3) -> 0, but the paper's own admissible tuning stated after Theorem 3.3 is tau ~ n^{-2/5} and s_n = o(n^{1/55}). For those rates, sqrt(n tau^3) = n^{-1/10}, so s_n / sqrt(n tau^3) = s_n n^{1/10} -> infinity, not zero. Since Theorem 3.4 invokes Theorem 3.2, no rate example actually satisfies all hypotheses as written. The proof of Theorem 3.2 never uses s_n / sqrt(n tau^3); it relies on bounds like s_n tau = o(1). So this is very likely a typo or an overly strong condition, but it is a genuine gap that a referee should force the authors to fix, either by correcting the condition or by supplying a compatible rate.\n\nTwo smaller issues. First, the arXiv abstract promises a real data analysis of energy consumptions, but the paper itself contains no such section and its own internal abstract mentions only simulations. That needs reconciling. Second, the simulation tuning uses s_n = floor((log n)^2/6), which for n <= 1000 gives s_n around 4 to 7, while the theory requires s_n / log n -> infinity. So the finite-sample settings are outside the proven regime, which is common in this literature but should be acknowledged.\n\nThe geometric physical-dependence decay (chi^i in Definition 2.1 and Assumption 3.1(2)) is strong, but it is explicit and satisfied by the paper's examples, so it is a scope restriction rather than a hidden flaw. No machine-checked proofs or pinned code are shipped; the GitHub link to an R package is a positive but unverifiable from the manuscript.\n\nThis paper is for researchers working on functional time series diagnostics, especially with multivariate non-stationary curve data. It deserves a serious referee, with the expectation of a revision to fix the Theorem 3.2 condition and align the abstract with the content.","headline":"Deserves a serious referee: the Gaussian approximation is a real advance, but Theorem 3.2's rate condition contradicts the paper's own tuning example and needs a correction.","tokens_in":69688,"tokens_out":4280,"would_cite":true,"duration_ms":38021,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M10","62R10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that multivariate locally stationary functional white noise can be tested without dimension reduction via a portmanteau statistic, whose distribution is handled by a new Gaussian approximation for degenerate U-statistics…","keywords":["multivariate functional time series","white noise testing","locally stationary processes","portmanteau test","high-dimensional Gaussian approximation","bootstrap","spatio-temporal data","U-statistics"],"falsifier":"Simulate a multivariate locally stationary functional process whose physical dependence measure decays polynomially rather than geometrically, compute the sup-norm distance in Theorem 3.1, or run the bootstrap test and check whether rejection rates approach the nominal level; a cheaper check is to compute the tail bound $\\Theta_{m,s}=\\sum_{l=m}^\\infty \\theta_{l,s}$ for the paper's linearized statistic and verify whether it is $O(\\chi^m)$, which fails under polynomial dependence.","tokens_in":68576,"feed_emoji":"📈","tokens_out":10677,"duration_ms":87485,"temperature":0.7,"pith_summary":"This paper establishes that the classic Box–Pierce portmanteau idea—test whether a time series is white noise by adding up estimated autocorrelations across lags—can be carried over to multivariate functional time series whose dynamics vary smoothly over time, without projecting the curves onto a finite-dimensional basis and with the number of lags growing with the sample size. The central claim is a Gaussian approximation: the sup-norm of the proposed statistic $Q_n/\\sqrt{s_n}$, a maximum over a time–argument grid of a sum of estimated squared Frobenius norms of autocovariances over $s_n$ lags, is uniformly close to the sup-norm of a single Gaussian vector that shares the autocovariance structure of the estimated linearized components. Because the limiting distribution may not exist, the paper proves that a difference-based block multiplier bootstrap yields correct asymptotic critical values, and that the resulting test is consistent against alternatives where the accumulated squared autocovariance over some set of lags is bounded away from zero. A reader should care because functional data in energy, environment, and finance are often multivariate, non-stationary, and high-dimensional, and existing white noise tests either assume stationarity or first compress the data by functional principal components, which can miss signals orthogonal to the chosen basis.","feed_headline":"New test detects white noise in time-varying functional series","feed_subtitle":"A bootstrap supplies critical values even when no limiting distribution exists, with lag count growing with sample size.","key_machinery":"The load-bearing object is the degenerate U-statistic representation of the portmanteau statistic. For each grid point, $\\sum_{k=1}^{s_n}\\operatorname{tr}\\{\\hat G_k(t,u)\\}$ is written as a U-statistic whose kernel $k_n(\\cdot,\\cdot,t)$ depends on the sample size; after replacing residuals by true errors and $m$-approximating the filter, the kernel becomes degenerate, which lets the statistic be linearized as $\\frac{1}{n\\tau s_n}\\sum_{i=1}^{2\\lceil n\\tau\\rceil}\\hat V_i$ with $\\hat V_i$ an $N(n-2\\lceil n\\tau\\rceil+1)$-dimensional vector of kernel-weighted traces. The distribution of this high-dimensional, locally stationary, increasingly dependent sum is then transferred to a Gaussian vector via an extension of high-dimensional Gaussian approximation (Proposition B.1), and the same $\\hat V_i$ vectors feed the difference-based block multiplier bootstrap that produces critical values.","core_discovery":"The paper's own claim is Theorem 3.1: under geometric physical-dependence and smoothness assumptions on the multivariate locally stationary functional filter, for the statistic $Q_n$ defined in (2.17) there exists a Gaussian vector $(Z_i)_{i=1}^{2\\lceil n\\tau\\rceil}$ with the autocovariance structure of $(V_i)$, such that $\\sup_x |P(Q_n/\\sqrt{s_n}\\le x)-P(|\\sum_{i=1}^{2\\lceil n\\tau\\rceil}Z_i|_\\infty/(s_n\\sqrt{n\\tau})\\le x)|=o(1)$. Theorems 3.3 and 3.4 then show the bootstrap quantile $\\tilde r_{\\rm boot}$ from Algorithm 1 makes the rejection rule $Q_n/\\sqrt{s_n}>\\tilde r_{\\rm boot}$ asymptotically level $\\alpha$ under the null of functional white noise and have power tending to 1 under alternatives (3.3). The test thereby detects serial dependence in multivariate functional data without stationarity or dimension reduction, accumulating dense signals over lags, dimensions, and time points.","pith_inferences":["The max-over-grid combined with sum-over-lags construction should also give the test power against sparse alternatives—signal concentrated in one lag, one dimension, or a narrow time interval—where sum-only portmanteau statistics dilute the evidence; verifying this on sparse alternatives is a natural next experiment.","The same Gaussian approximation machinery could produce simultaneous confidence bands for the autocovariance operator of a locally stationary functional series, turning the test into a full inference tool.","The geometric-decay assumption is likely stronger than needed in practice—the simulation models are only moderately dependent—so an extension to polynomially decaying physical dependence, possibly at the cost of slower rates, is a plausible direction the paper leaves open."],"forward_implications":["The test provides a white-noise diagnostic for multivariate functional data with time-varying data-generating mechanisms, so practitioners no longer need to assume stationarity before checking residual assumptions.","Because it does not rely on a Karhunen–Loève expansion, it detects signals that are orthogonal to any chosen finite-dimensional basis, overcoming a known limitation of FPCA-based portmanteau tests.","The number of lags $s_n$ can grow with $n$; with the paper's rate choices, e.g. $s_n=o(n^{1/55})$, the statistic pools evidence over many lags while the bootstrap still controls the level.","The Gaussian approximation for degenerate second-order U-statistics of locally stationary functional time series is a standalone result that can be reused for other test statistics with the same structure.","In simulations the nominal level is well approximated from $n=400$ upward and power increases with sample size, dimension, and signal strength $\\delta$ across functional AR and time-varying functional GARCH alternatives."],"supporting_citations":[{"why":"Supplies the high-dimensional Gaussian approximation for dependent vectors that Proposition B.1 extends to increasing dependence and finite moments.","marker":"Zhang and Cheng (2018)"},{"why":"Provides the Gaussian-to-Gaussian comparison and multiplier bootstrap error bounds used inside the proof of Proposition B.1.","marker":"Chernozhukov, Chetverikov and Kato (2013)"},{"why":"Gives kernel-estimation and local-stationarity lemmas used to control estimation error of the mean and autocovariances.","marker":"Dette, Wu and Zhou (2019)"},{"why":"Supplies Lemma S1 and moving-sum variance estimates used to transfer the Gaussian approximation to the bootstrap statistic.","marker":"Dette and Wu (2026)"},{"why":"Provides Lemma 5 on the decay of autocovariances and the physical-dependence chaining arguments underlying the cumulant bounds.","marker":"Zhou and Wu (2010)"},{"why":"Establishes existence and geometric dependence bounds for iterated random functions used in Example 2.2 and Remark 3.1.","marker":"Wu and Shao (2004)"},{"why":"Is the closest predecessor: a portmanteau test for univariate locally stationary functional time series that this paper extends to the multivariate setting.","marker":"Bücher, Dette and Heinrichs (2023)"},{"why":"Baseline portmanteau test for stationary functional time series under conditional heteroscedasticity; the paper contrasts its FPCA-based approach and avoids dimension reduction.","marker":"Kokoszka, Rice and Shang (2017)"},{"why":"Another stationary functional portmanteau-type baseline whose sum-of-L2-norms statistic the new max-plus-sum statistic generalizes.","marker":"Rice, Wirjanto and Zhao (2020)"},{"why":"Shows vector-valued martingales cannot generally be embedded in a Gaussian process, motivating the direct Gaussian approximation instead of a martingale CLT.","marker":"Monrad and Philipp (1991)"}],"fun_headline_variants":["Bootstrap portmanteau test for non-stationary functional series","Test for white noise in time-varying functional data without dimension reduction","Increasing-lag portmanteau test handles non-stationary functional series","No-stationarity test sniffs out dependence in multivariate functional data","Portmanteau test with many lags for non-stationary functional time series"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument presumes that each past innovation's influence on the functional filter, and on its derivative, decays geometrically fast; if that influence decays only polynomially, the cumulant bounds controlling the U-statistic's effective dimension cease to hold and the Gaussian approximation is no longer proven.","fun_headline_variants_meta":{"raw":{"variants":["Bootstrap portmanteau test for non-stationary functional series","Test for white noise in time-varying functional data without dimension reduction","Increasing-lag portmanteau test handles non-stationary functional series","No-stationarity test sniffs out dependence in multivariate functional data","Portmanteau test with many lags for non-stationary functional time series"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"total_tokens":1268,"prompt_tokens":942,"completion_tokens":326,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":229}},"tokens_in":558,"tokens_out":326,"duration_ms":3594,"temperature":1.0,"reasoning_tokens":229,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:00:12.465020+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a multivariate locally stationary functional process whose physical dependence measure decays polynomially rather than geometrically, compute the sup-norm distance in Theorem 3.1, or run the bootstrap test and check whether rejection rates approach the nominal level; a cheaper check is to compute the tail bound $\\Theta_{m,s}=\\sum_{l=m}^\\infty \\theta_{l,s}$ for the paper's linearized statistic and verify whether it is $O(\\chi^m)$, which fails under polynomial dependence.","supporting_citations":[{"cited_title":", Wu , Weichi W","cited_arxiv_id":null,"evidence_quote":"Gives kernel-estimation and local-stationarity lemmas used to control estimation error of the mean and autocovariances."},{"cited_title":"Confidence surfaces for the mean of locally stationary functional time series","cited_arxiv_id":"2109.03641","evidence_quote":"Supplies Lemma S1 and moving-sum variance estimates used to transfer the Gaussian approximation to the bootstrap statistic."},{"cited_title":"Wu , Wei Biao W","cited_arxiv_id":null,"evidence_quote":"Provides Lemma 5 on the decay of autocovariances and the physical-dependence chaining arguments underlying the cumulant bounds."},{"cited_title":", Wirjanto , Tony T","cited_arxiv_id":null,"evidence_quote":"Another stationary functional portmanteau-type baseline whose sum-of-L2-norms statistic the new max-plus-sum statistic generalizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows vector-valued martingales cannot generally be embedded in a Gaussian process, motivating the direct Gaussian approximation instead of a martingale CLT."}],"review_version":1}