{"id":"0b343726-1130-4d52-ac70-6b45c2f24b60","arxiv_id":"2607.20828","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Adaptive change-point tests combining quadratic and max CUSUM statistics are extended to non-Gaussian temporally dependent high-dimensional data, with validated limiting distributions and multi-change recovery by wild binary segmentation.","lead":"This paper builds statistical tests and estimators for sudden level shifts in high-dimensional time series where observations are serially dependent. A generalist should care because these tools handle both widespread and concentrated changes in one framework, with applications to macroeconomic panels, electricity demand, and sensor data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assumption 3 is not implied by Assumptions 1–2; a simple iid scale-mixture process satisfying Assumptions 1, 2, and 4 has C_{pi,p} = O(p^2), so the claimed Gaussian limit and matched independence for 'general non-Gaussian' dependence are unsupported.","rationale":"Both the reader and I identify Assumption 3 as the weakest load-bearing point. I went further and found an explicit process within Assumptions 1, 2, and 4 that violates it, showing it is not a vacuous technical condition. The paper's simulation section does not help: the Gaussian matrix filter has vanishing connected cumulants, and the Gamma design is explicitly outside Assumption 1. The rate contradiction in Theorem 2 and the log n = o(p^{1/4}) condition are real but secondary; the former is likely a typo (M/sqrt(n) from the proof) and the latter concerns the gamma=1/2 regime's asymptotic condition, not the logical validity of the theorems. The load-bearing failure mode is the quadratic path's Gaussian comparison, which collapses if connected cumulants are superlinear in p. The proposed check uses a simple iid scale-mixture process to test whether the null distribution of S_0 actually matches F_V; this settles whether the concern lands. Since the theorems remain valid conditional on Assumption 3, and the paper is explicit about its assumptions, I recommend no change to the CONDITIONAL verdict.","tokens_in":91161,"tokens_out":14436,"duration_ms":148506,"concrete_test":"Analytic check plus simulation: for the process eps_{t,j}=f_t eta_{t,j} with f_t iid uniform[-√3,√3], eta_{t,j} iid Rademacher, verify in closed form that the q=2 connected cumulant sum equals 0.8 p^2 (so Assumption 3 fails). Then simulate n=1000, p=500 under H0, compute S_0 with the paper's feasible centering/scale (§2), and compare the empirical distribution of S_0 to F_V over 10^4 replications (e.g., 95% quantile vs the paper's b_c0.95 and a KS test). If the empirical null distribution deviates by more than simulation error, Theorem 1's Gaussian limit and Theorem 5's independence fail for an in-scope process; if it matches despite the violated assumption, the proof of Assumption 3's necessity is incomplete.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Assumption 3 is the load-bearing hinge: Lemmas S.3–S.5, S.19, and S.24 all invoke it to control fourth-, sixth-, and eighth-order connected cumulants of the quadratic CUSUM, so Theorem 1's Gaussian-process limit and Theorem 5's matched independence (hence the Cauchy combination) are valid only if it holds. The paper gives no primitive sufficient conditions and no in-assumption non-Gaussian simulation (the Gamma design is outside Assumption 1). More seriously, Assumption 3 can fail inside the stated framework. Let f_t be iid uniform on [-√3,√3], eta_{t,j} iid Rademacher, and set eps_{t,j}=f_t eta_{t,j}. This process is m=0, projection sub-Gaussian, Omega=I_p, tr(f_p(lambda))/p=1/(2pi), and Assumption 4 holds. But for the connected single-block partition in q=2, C_{pi,p}=sum_{j1,j2}(E f^4 - 1)=0.8 p^2 = O(p^2), violating Assumption 3. Thus the central theorem's hypotheses are not satisfied by a simple temporally independent, cross-sectionally dependent sub-Gaussian process, and the claim of covering 'general non-Gaussian' dependence is not supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes sparsity-adaptive tests and localizers for a high-dimensional mean change under temporal dependence. It pairs a quadratic CUSUM scan (dense signal) with a coordinatewise maximum CUSUM scan (sparse signal), under two temporal weightings: an unweighted full scan and a variance-standardized trimmed scan. For each statistic the authors introduce dependence-adjusted centering and long-run scale estimation, prove Gaussian-process or Gumbel null limits, and prove asymptotic independence of the matched quadratic and maximum statistics, which justifies Cauchy combination tests. They also prove single-change localization rates and consistent multiple-change recovery via guarded wild binary segmentation. The paper includes an extensive supplement, simulations for Gaussian and standardized Gamma innovations, and applications to FRED-MD and electricity-load panels.","tokens_in":91494,"tokens_out":7334,"duration_ms":76815,"significance":"If correct, the paper would fill a real gap: joint null calibration of dense and sparse CUSUM scans under non-Gaussian temporal dependence, plus localization and multiple-change recovery within one framework. The supplement is unusually detailed and the simulations are honestly reported, including the explicit concession that the Gamma design lies outside Assumption 1 and the post-hoc white-noise diagnostics showing remaining dependence after mean removal. However, the central \"general non-Gaussian\" claim rests on Assumption 3, which is not implied by Assumptions 1–2 and fails for a simple temporally independent sub-Gaussian process. In addition, the rate in Theorem 2 contains an impossible term, and the boundary-weighted results are only proved under a growth condition violated by every simulation and application. These issues are load-bearing; the manuscript needs substantial revision before its main claims can be accepted.","major_comments":[{"comment":"Assumption 3 is the hinge for the Gaussian comparison steps that produce Theorem 1, Theorem 5, and the WBS theorems, but it is not a consequence of Assumptions 1–2 and it can fail inside the stated framework. Let f_t be iid uniform on [-sqrt(3), sqrt(3)], let eta_{t,j} be iid Rademacher, and set eps_{t,j}=f_t eta_{t,j}. This process is m=0-dependent, projection sub-Gaussian, has Omega=I, and satisfies Assumption 4. For the connected single-block partition in q=2, the blockwise cumulant envelope gives C_{pi,p} = (3/4)p(p-1) - (6/5)p = (3/4)p^2 - (27/20)p = O(p^2), violating Assumption 3. Thus the central theorem's hypotheses are not satisfied by a simple iid sub-Gaussian process, and the claim of covering \"general non-Gaussian\" dependence is unsupported. The paper gives no primitive sufficient conditions for Assumption 3, and the simulations provide no non-Gaussian evidence: the Gaussian","section":"Assumption 3 and Lemmas S.3–S.5, S.19, S.24"},{"comment":"The theorem defines r_mu,n = M*sqrt(n) + sqrt(p) M eta_M + ... and asserts r_mu,n -> 0. With M = ceil((n ^ p)^{1/8}), the leading term M*sqrt(n) diverges for every sequence satisfying p = o(n^{3/2}): if p >= n it is at least n^{5/8}, and if p is fixed it is order n^{1/2}. Therefore the stated uniform centering rate cannot be negligible, and the subsequent uses of r_mu,n in Theorem 3 and Theorem 5 are not justified. The proof of Lemma S.10 obtains a stochastic contribution of order M/sqrt(n), so it appears that M/sqrt(n) was intended. If so, correct Theorem 2 and re-verify the feasibility conditions in Theorems 3 and 5; otherwise prove the displayed rate or revise the statement.","section":"Theorem 2, definition of r_mu,n"},{"comment":"The boundary-weighted asymptotic statements require log n = o(p^{1/4}). This condition is violated in every simulation and in the FRED-MD application: for n=789, p=119, log n is about 6.67 while p^{1/4} is about 3.30; similarly, n=400, p=250 gives log n ~ 5.99 versus p^{1/4} ~ 3.98, and n=800, p=500 gives 6.68 versus 4.73. Since the boundary-weighted statistics S_{1/2}, M_{1/2}, and T_{CC,1/2} are used and reported as decisive in the applications, the paper is exercising its main calibration tools outside the proven regime. Please either add simulations and applications that satisfy the stated growth condition, or establish the boundary-weighted limits under less restrictive conditions.","section":"Theorem 3(ii), Theorem 5, and Section 6/7.1"},{"comment":"The only non-Gaussian simulation design is the standardized Gamma innovation, which the paper itself places outside Assumption 1. Thus no simulation demonstrates that Assumption 3 is satisfiable by a non-Gaussian process or that the Gaussian comparison bounds hold beyond the Gaussian-filter case. The empirical size and power results are informative as finite-sample illustrations, but they should not be read as evidence for the general non-Gaussian theorem. Please provide a non-Gaussian simulation within Assumptions 1–3, or state clearly that Assumption 3 is not numerically verified.","section":"Simulation Section 6: non-Gaussian coverage of Assumption 3"}],"minor_comments":[{"comment":"The term M*sqrt(n) in r_mu,n appears to be a typo for M/sqrt(n). The proof and the claimed convergence suggest the latter; please fix the display and check all places where r_mu,n is used.","section":"Theorem 2 notation"},{"comment":"The statement that the standardized Gamma design is outside Assumption 1 is honest, but the text later says the procedures \"retain their sparsity-adaptive behavior under the asymmetric non-Gaussian design.\" That is a finite-sample claim, not a validation of Theorems 1–5; please make this distinction explicit in the simulation discussion.","section":"Sections 2 and 6: Gamma design disclaimer"},{"comment":"The RCP preprocessing and finite-LRV tail formulas are described as finite-sample calibration choices not covered by the formal theorems, yet the reported p-values in the applications use them. This is acceptable if clearly labeled, but please add a sentence reminding the reader that the formal theorems apply to the asymptotic p-values, not to the implemented p_FLRV versions.","section":"Section 3 and Section 7: calibration implementation"},{"comment":"The supplement is very long and dense, with many locally defined objects. Adding a short table of notation, especially for the many superscripts on Gaussian analogues (M^{G,gamma}, M^{B,gamma}, M^{eps,gamma}) and for the multiple-change rates, would substantially improve readability.","section":"General presentation"}],"recommendation":"major_revision","confidential_remarks":"The manuscript contains a substantial and sophisticated proof supplement, and the empirical work is transparent. The main obstacle is not the length or style but the status of Assumption 3: it is high-level, not implied by the other assumptions, and demonstrably fails for a simple iid sub-Gaussian process. Unless the authors supply primitive sufficient conditions or substantially narrow the scope, the central claim of general non-Gaussian validity will remain unsupported. The rate error in Theorem 2 and the growth-condition violation in the simulations are also blocking issues."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nTwo things to know before spending time on this paper. It is the most complete attempt I know to unify dense and sparse change-point testing, single-change localization, and wild binary segmentation for high-dimensional time series under temporal dependence. It also rests on a high-level cumulant condition that is not implied by the other assumptions and is violated by a very simple sub-Gaussian process, so the abstract's 'general non-Gaussian' claim is not established.\n\nWhat is genuinely new: the edge-corrected centering with two trace orientations for nonsymmetric lag covariance matrices, the uniform rates for feasible centering and long-run scale, the asymptotic independence between matched quadratic and coordinatewise-max CUSUM statistics, and the WBS consistency theorems. The supplement is long, serious, and mostly coherent. Conditional on Assumption 3, the proof chain looks plausible. The simulations are honestly reported, including the admission that the Gamma design lies outside Assumption 1, and the size/power comparisons are informative.\n\nThe main soft spot is Assumption 3. It controls connected cumulant contractions, and essentially every Gaussian-comparison step for the dense process uses it. The paper gives no primitive sufficient conditions. This is not merely a verification gap: take eps_tj = f_t eta_tj with f_t iid uniform on [-sqrt3,sqrt3] and eta_tj iid Rademacher. That process is m-0-dependent, projection sub-Gaussian, has Omega = I_p, and satisfies Assumptions 1, 2, and 4. But for q=3 the connected single-block contraction is O(p^3), not O(p), so Assumption 3 fails. The stress-test note's q=2 calculation is off -- the fourth-order cumulant there is O(p), not O(p^2) -- but its conclusion is correct at order six. The paper needs primitive sufficient conditions for Assumption 3, or a substantial narrowing of scope.\n\nSmaller issues: Theorem 2 displays r_mu,n = M*sqrt(n), which diverges; the proof implies M/sqrt(n), so it is likely a typo but one that should not survive. The condition log n = o(p^{1/4}) is violated in every numerical experiment, so the boundary-weighted calibration is exercised outside the proven regime. WBS tuning for the empirical applications is not reported, and no code is shipped.\n\nThis paper is for specialists in high-dimensional change points, especially those extending sparsity-adaptive methods to dependent data. It deserves a serious referee: the framework is useful and the supplement is extensive, but the referee letter should center on Assumption 3 and demand primitive conditions plus an in-assumption non-Gaussian simulation. I would send it out, not desk-reject.","headline":"Real and substantial framework for sparsity-adaptive change-point inference under temporal dependence, but the load-bearing Assumption 3 can fail for a simple sub-Gaussian process, so the advertised 'general non-Gaussian' coverage is not supported.","tokens_in":92011,"tokens_out":8688,"would_cite":true,"duration_ms":93239,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M10","62H15","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that high-dimensional mean-change tests and localization remain valid under temporal dependence by pairing quadratic and coordinatewise-maximum CUSUM scans.","keywords":["change point detection","high-dimensional time series","CUSUM statistics","Cauchy combination test","long-run covariance","wild binary segmentation","temporal dependence","sparsity adaptation"],"falsifier":"Simulate a stationary non-Gaussian vector MA(2) with heavy-tailed innovations at n=800, p=250, estimate the connected cumulant sums in Assumption 3 from the residuals, and compare the empirical null distribution of S0 and S_{1/2} with the claimed F_V and Gumbel limits. If those cumulant sums are not O(p), or if the empirical sizes deviate from the nominal level once the assumed growth rates (e.g., p >> (log n)^4) are violated, the central claim fails.","tokens_in":90973,"feed_emoji":"📊","tokens_out":7031,"duration_ms":62493,"temperature":0.7,"pith_summary":"The paper sets out to prove that detecting and locating a mean change in a high-dimensional time series does not require temporal independence or Gaussianity. It pairs two scans — a quadratic CUSUM that aggregates evidence across all coordinates (for dense changes) and a coordinatewise maximum CUSUM (for sparse changes) — and recalibrates both for serial dependence using long-run covariance estimates and dependence-adjusted centering. Its central result is that matched quadratic and maximum scans are asymptotically independent under the null, so combining them through a Cauchy test preserves size while adapting to unknown sparsity. The paper also proves consistency and rates for single-change localization and, via guarded wild binary segmentation, consistent recovery of multiple changes. If correct, this gives practitioners one set of procedures for testing, dating, and enumerating changes in dependent high-dimensional panels.","feed_headline":"Dependent high-dim series get valid change-point tests","feed_subtitle":"Matched dense and sparse CUSUM scans stay independent under dependence, enabling adaptive detection and localization.","key_machinery":"The load-bearing object is the matched pair of CUSUM scans: the quadratic statistic W(k), centered by a dependence-adjusted mean and scaled by a long-run scale that keeps the two trace orientations tr{Gamma(h)Gamma(k)^T} and tr{Gamma(h)Gamma(k)} separate; and the coordinatewise maximum M, standardized by difference-based long-run marginal variance estimates. Two temporal weightings (gamma=0 unweighted, gamma=1/2 variance-standardized) are paired. The asymptotic independence of a matched pair — with the F_V supremum limit for the unweighted quadratic path and Gumbel limits for the max and boundary-weighted paths — is the mechanism that legitimizes the Cauchy combination. Wild binary segmentat","core_discovery":"The paper's central claim is that, under non-Gaussian vector dependence with sub-Gaussian projections and a higher-order connected-cumulant contraction condition, the feasible quadratic CUSUM path converges to a Gaussian process with supremum distribution F_V, the boundary-weighted quadratic and both maximum scans converge to Gumbel limits after Darling–Erdős normalization, and matched quadratic/maximum pairs are asymptotically independent — for instance Pr{S0 <= x, 2M0^2 − log(2p) <= y} -> F_V(x)F_Gumbel(y). Those independence facts justify Cauchy-combination tests that adapt to dense or sparse alternatives. The same centered scores localize a single change with explicit rates, and a guarde","pith_inferences":["As a reader inference, the boundary-weighted calibration is proven only when p grows faster than (log n)^4; with n=789 and p=119 (the macro application), log n ≈ 6.7 exceeds p^{1/4} ≈ 3.3, so that reported accuracy is an empirical demonstration outside the proven regime rather than a theorem consequence.","As a reader inference, the finite-sample Gamma tail corrections and preprocessing used in the implementation are not invoked in the formal proofs; comparing them to the asymptotic Gumbel calibration at small p would directly measure how much of the size control is proof versus calibration.","As a reader inference, the dense and sparse components often select different dates in the empirical panels, suggesting the matched scans can be read as a decomposition of a change into broad shifts and concentrated shifts — a use the paper mentions but does not formalize.","As a reader inference, a natural extension, pointed to by the paper's own outlook, is a diverging number of changes with shrinking spacing; the fixed-proportion spacing assumption in the WBS theorems is what keeps the shortest-significant selection valid."],"forward_implications":["Cauchy-combined tests are asymptotically size-valid under serial dependence for both dense and sparse alternatives, without knowing sparsity in advance.","For a detected change, the adaptive estimator chooses the dense or sparse localizer by component p-values and is consistent whenever one component is consistent.","Guarded wild binary segmentation consistently estimates the number of change points and attains localization rates, with estimated locations converging relative to the minimum spacing.","The variance-standardized boundary-weighted scans have a Gumbel null limit, so near-boundary changes can be detected and localized without simulating a new reference distribution.","Ignoring serial dependence in these settings is not harmless: the paper's simulations show independence-calibrated procedures reject in nearly every replication under an MA(2) design."],"fun_headline_variants":["Cauchy combo detects dense or sparse shifts in high-dim","High-dim change points: adaptive dense-sparse tests","Dense or sparse? New test catches both in high-dim","High-dim change detection adapts to dense and sparse"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is Assumption 3: sums of higher-order temporal cumulant blocks must grow only linearly in dimension, and no primitive data-checkable condition is given for it (with a secondary constraint that boundary-weighted theorems need p growing faster than (log n)^4, which the paper's simulations do not meet).","fun_headline_variants_meta":{"raw":{"variants":["Cauchy combo detects dense or sparse shifts in high-dim","High-dim change points: adaptive dense-sparse tests","Dense or sparse? New test catches both in high-dim","High-dim change detection adapts to dense and sparse"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001182,"raw_usage":{"total_tokens":4656,"prompt_tokens":617,"completion_tokens":4039,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":361,"completion_tokens_details":{"reasoning_tokens":3973}},"tokens_in":361,"tokens_out":4039,"duration_ms":23638,"temperature":1.0,"reasoning_tokens":3973,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T09:17:38.767848+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a stationary non-Gaussian vector MA(2) with heavy-tailed innovations at n=800, p=250, estimate the connected cumulant sums in Assumption 3 from the residuals, and compare the empirical null distribution of S0 and S_{1/2} with the claimed F_V and Gumbel limits. If those cumulant sums are not O(p), or if the empirical sizes deviate from the nominal level once the assumed growth rates (e.g., p >> (log n)^4) are violated, the central claim fails.","supporting_citations":[],"review_version":1}