{"id":"0decf1c7-0396-4502-9a4c-86630f06d56e","arxiv_id":"2607.22162","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ERHT is a new robust, dependence-aware high-dimensional change-point test that is asymptotically calibrated through Gaussian-process limits and consistently localizes multiple breaks by wild binary segmentation.","lead":"This paper introduces ERHT, a change-point test for high-dimensional data that combines robust spatial medians with ridge-regularized dependence weighting, and proves asymptotically exact Gaussian-process calibrations for it. If correct, it gives analysts a calibrated tool for finding structural breaks in fat-tailed, strongly correlated data such as industry stock returns.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: the t3-moment concern rests on a sign error; chi-square positive moments are all finite.","rationale":"The reader's central concern—that Assumption 3.1 fails for multivariate t_3 because chi-square moments only exist below order 1.5—is a mathematical mistake. For S_ν ~ χ^2_ν, positive moments of every order exist; only negative (inverse) moments are constrained by the density near zero. Since ξ involves (S_ν/ν)^{1/2} to a positive power, Eξ^{4+η} is finite for ν=3. The paper's S1.1 verification is correct. After correcting this, I examined the rest of the argument. The main theorems rest on detailed and internally consistent lemmas (Rademacher sign representation, deterministic equivalents, covariance factorization, tightness). The paper is transparent about the analytic Cauchy rule's non-exactness and the distinction from joint-limit calibration; the exact calibration claim is confined to the joint-limit procedure. The reliance on unpublished sibling papers and lack of code are practical limitations, not logical flaws. Therefore I find no load-bearing concern that changes the reader's verdict; the CONDITIONAL status can remain for secondary reasons (implementation gaps), but the specific t_3 objection should be withdrawn.","tokens_in":73954,"tokens_out":25781,"duration_ms":245272,"concrete_test":"Directly compute E(S_3/3)^{(4+η)/2} for η=0.1 via the Gamma formula: it equals 3^{-(4+η)/2} 2^{(4+η)/2} Γ((4+η)/2+3/2)/Γ(3/2), which is finite, confirming S1.1. As a secondary check, simulate the null distribution for t_3 errors with p=100, n=200 and verify that the empirical covariance of Z_ρ(s_t) and Z_ρ(s_u) approaches K_0(t,u) as the sample size grows.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption is not valid. For S_ν ~ χ^2_ν, E(S_ν)^t = 2^t Γ(t+ν/2)/Γ(ν/2) converges for every t > -ν/2; in particular all positive t are allowed. The radial variable ξ = c_ν^{-1} C_p (S_ν/ν)^{1/2} has E ξ^{4+η} proportional to E(S_ν)^{(4+η)/2}, which is finite for ν=3. The constraint E D_ν^r < ∞ only for r<ν applies to negative (or inverse) moments of D_ν, not positive ones. Thus Assumption 3.1 is satisfied by the t_3 design, and S1.1's verification is correct. I find no other internally inconsistent or unsupported step that threatens the central Gaussian-process limits, the joint convergence over the ridge grid, or the WBS consistency. The analytic Cauchy rule is explicitly characterized rather than asserted to be exactly uniform, and the exact joint-limit calibration is a theoretical option; neither is a logical flaw in the theorems.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an elliptical regularized Hotelling (ERHT) procedure for high-dimensional location change-point detection under heavy-tailed, cross-sectionally dependent elliptical observations. The statistic contrasts spatial medians of adjacent segments and studentizes the contrast using a ridge-regularized inverse of the pooled centered spatial-sign covariance matrix. The main theoretical results are: (i) uniform raw-to-score reduction and pointwise null laws (Prop. 3.1, Thm. 3.1); (ii) Gaussian-process limits for the single- and multiple-change scan statistics with covariance kernel K0(s,r)=ψ(s,r)^2/(ψ(s)ψ(r)) (Thms. 3.2, 3.7); (iii) joint convergence over a finite ridge grid with cross-parameter correlation r_E(ρ,ρ') (Thms. 3.3, 3.8), yielding exact asymptotic calibration of the Cauchy-aggregated test (Thms. 3.4, 3.9); (iv) local power and localization rates for single-change alternatives (Thms. 3.5, 3.6); and (v) WBS-ERHT consistency for estimating the number and locations of multiple changes (Thm. 4.1). Simulations compare ERHT with covariance-based RHT, mean-based DMS0, and spatial-sign SSCPD0 across normal, t3, and Gaussian-mixture errors; a Fama–French 49 industry portfolio analysis reports four structural breaks.","tokens_in":74191,"tokens_out":29049,"duration_ms":321251,"significance":"If the theorem chain holds, the paper makes a substantial contribution: it extends regularized Hotelling methodology to heavy-tailed elliptical data in the proportional-growth regime, provides explicit Gaussian-process covariance kernels and joint-limit calibration, and gives a complete WBS consistency theory with localization rates. The strengths are the detailed supplementary proof chain, the use of deterministic equivalents and companion laws derived from first principles, and the honest distinction between exact joint-limit calibration and the analytic Cauchy rule. I specifically checked the concern that the t3 simulation design violates Assumption 3.1. That concern is based on a sign error: for S_ν ~ χ²_ν, E(S_ν)^t is finite for every positive t, so Eξ^{4+η} is finite for the multivariate t3 construction; the verification in S1.1 is correct. The main practical caveat is that the analytic Cauchy transformation used in the numerical sections is characterized but not proven to control the asymptotic level; this does not undermine the central Gaussian-process or WBS theorems, but it should be stated more prominently.","major_comments":[],"minor_comments":[{"comment":"The numerical method called ERHT-CC in Sections 5 and 6 uses the analytic Cauchy transformation, but Theorems 3.4 and 3.9 only characterize its limiting rejection probability as P{T_∞ ≥ cot(πα)}; they do not establish that this is ≤ α under the dependence among the P_k's. The exact joint-limit calibration is a theoretical option but is not implemented in the experiments. Please state this limitation explicitly where ERHT-CC is defined and used, so that readers do not treat the analytic P_CC as an exact p-value.","section":"Section 2.3 and Theorems 3.4/3.9"},{"comment":"The statement that subtracting the coordinatewise full-sample mean 'has no effect on either segment contrasts or centered spatial signs' is imprecise: subtracting a random vector changes the origin used for spatial signs, and coordinatewise demeaning by a data-dependent mean can break the exact elliptical symmetry assumed by the model. Please rephrase to say that the mean shift does not affect the difference between segment mean contrasts, or provide a formal justification for the preprocessing within the elliptical model.","section":"Section 6.1"},{"comment":"The notation for the maximum and minimum jump signal, s_WBS and ̲s_WBS, is easy to confuse in both Eq. (24) and Assumption 4.2. Please use distinct symbols consistently throughout the statements and proofs.","section":"Section 4.2, Eq. (24) and Assumption 4.2"},{"comment":"Some ERHT-CC empirical sizes are noticeably above the nominal 5% level, e.g., Identity/Normal/p=400/n=200 shows 8.1%. A brief comment on finite-sample calibration, or additional Monte Carlo replications, would help the reader assess whether this is sampling noise or a small-sample bias.","section":"Table 1"},{"comment":"The panel label 'Ploy' appears in multiple-change power figures; it should read 'Poly'.","section":"Figures 4-6"},{"comment":"Please include a data/code availability statement. The Fama–French data are public, but the exact preprocessing steps, permutation scheme, and tuning choices should be documented for reproducibility.","section":"General"}],"recommendation":"minor_revision","confidential_remarks":"I found no internally inconsistent proof step that threatens the central theorems. The t3 moment concern raised in the stress-test note does not land: S1.1's verification is correct. The more substantive practical caveat is that the analytic Cauchy rule used in the simulations lacks an asymptotic level guarantee; I regard this as a presentation/validation issue rather than a reason to reject. The paper would benefit from making the heuristic status of the analytic rule explicit and, ideally, from reporting exact joint-limit calibration or a bound on its rejection probability."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the paper, the reader's report, and the stress-test note. The reader's central objection is wrong, and the stress-test note is right. For S ~ chi^2_3, all positive moments exist; E(S^t) is finite for every t > -3/2. So the t_3 design satisfies Assumption 3.1, and the S1.1 verification is correct. The confusion comes from applying the negative-moment restriction to positive orders. Drop that concern.\n\nWhat is actually new: ERHT combines a spatial-median contrast with a ridge-regularized inverse of the centered spatial-sign covariance matrix, and the theory is genuinely derived rather than fitted. The Gaussian-process limits for the scan statistics (Theorems 3.2 and 3.7), the joint convergence over the ridge grid with the r_E correlation, the local power and localization rates, and the WBS consistency theorem form a long but internally coherent chain. I spot-checked the key structural steps — raw-to-score reduction, Rademacher companion representation, the covariance factorization, tightness — and they hold together. The paper is also honest: it explicitly distinguishes the analytic Cauchy rule from exact joint-limit calibration, acknowledges the exchangeability assumption, and describes the real-data break associations as descriptive.\n\nThe real soft spots are minor and mostly practical. The analytic Cauchy aggregate is not exactly calibrated at level alpha; the paper characterizes its limiting rejection probability but does not provide exact finite-sample size control. The theoretically exact joint-limit calibration depends on the limiting spectrum through r_E, which is never implemented. Some foundational inputs are unpublished sibling preprints, so a referee will have to trust or verify those. No code is shipped, which makes the simulations harder to reproduce. None of these undercut the central theorems; they mean the paper is theory-first and the default recipe is the permutation-based version.\n\nFor whom: this is for statisticians working on high-dimensional change-point detection, especially those who care about heavy tails with cross-sectional dependence. It deserves a serious referee — the conditional verdict should become accept with minor-to-moderate revision after the moment misreading is cleared up. I would bring it to a reading group and would cite it in my own work if I publish in this area.","headline":"ERHT is a substantial, carefully derived contribution to robust high-dimensional change-point testing; the flagged t_3 moment issue is a false alarm, and the real caveats are implementation-level rather than mathematical.","tokens_in":74723,"tokens_out":1756,"would_cite":true,"duration_ms":21054,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H15","62G10","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A robust scan statistic for high-dimensional change-point detection is shown to converge to a Gaussian process with covariance kernel K0(s,r)=ψ(s,r)^2/(ψ(s)ψ(r)), enabling asymptotically exact calibration and consistent multiple-break local","keywords":["change-point detection","high-dimensional location testing","elliptical distribution","spatial median","spatial-sign covariance","ridge regularization","Gaussian-process limit","Cauchy aggregation"],"falsifier":"Simulate the ERHT single-change scan under multivariate t_3 errors with p/n≈γ and no change, compute Gaussian-supremum-calibrated p-values from F_sc, and check empirical size: if the convergence theorems were valid under t_3, rejection rates would approach α; if, as the moment gap suggests, rates deviate systematically as n,p grow, the theorems' hypotheses are not met. Equivalently, compute E(R^{-1})^{4+η} for t_3 to see it is infinite for every η>0.","tokens_in":73700,"feed_emoji":"📊","tokens_out":5036,"duration_ms":54111,"temperature":0.7,"pith_summary":"The paper proposes ERHT, a two-sample scan test for location changes in high-dimensional elliptical sequences. It contrasts spatial medians of adjacent segments and standardizes the contrast by a ridge-regularized inverse of the pooled centered spatial-sign covariance matrix. The main claim is that the studentized scan process converges to a centered Gaussian process whose covariance kernel is a squared overlap of temporal contrast functions, with joint convergence across a grid of ridge parameters. That limit yields asymptotically exact critical values via the supremum distribution and a Cauchy aggregation rule, plus local power and localization rates, and a wild-binary-segmentation extension that consistently estimates the number and locations of multiple breaks. If correct, the method gives valid p-values in the heavy-tailed, cross-sectionally dependent high-dimensional regime.","feed_headline":"Robust high-dimensional change-point scans have a Gaussian limit","feed_subtitle":"Spatial medians plus ridge-weighted shape normalization give exact calibration and break localization for heavy-tailed high-dimensional data","key_machinery":"The companion-matrix representation V_raw ≈ m β^⊤ A_ρ β combined with a Rademacher sign representation reduces the nonlinear spatial-median contrast to a conditionally quadratic form in independent signs. This gives explicit centering κ and variance σ², and the deterministic-equivalent resolvent D_ρ,n = (a_ρ,n Ω_p + ρI)^{-1} from the Marchenko-Pastur fixed point transfers the cross-sectional geometry to the covariance kernel K0(s,r)=ψ(s,r)²/(ψ(s)ψ(r)) and the cross-ridge correlation r_E(ρ,ρ').","core_discovery":"For each candidate scan interval s=(t1,t2,t3), the statistic Z_ρ(s) compares spatial medians of two adjacent windows after studentization by a ridge-regularized inverse of the spatial-sign covariance matrix. The paper proves that over the single-change scan {(0,t,1)} and the multiple-change scan {t1<t2<t3} the process {Z_ρ(s)} converges to a centered Gaussian process with covariance K0(s,r)=ψ(s,r)^2/(ψ(s)ψ(r)), where ψ(s,r) is the L^2 overlap of the two temporal contrast functions. Joint convergence over the ridge grid holds with cross-parameter correlation r_E(ρ,ρ'), and because the marginal law is ρ-independent under the common full-sample pool, the same supremum calibration applies at eve","pith_inferences":["Because the common-pool assumption makes the marginal Gaussian law independent of ρ, the paper's calibration step implicitly turns the choice of ridge parameter into a nuisance; one could extend the same aggregation to a continuum of ρ values, or to maximum-type combinations, with the same limit theory.","The radial moment condition E ξ^{4+η}<∞ is the point most likely to fail in practice: for multivariate t_ν, E ξ^r is finite only when r<ν, so the showcased t_3 design does not satisfy Assumption 3.1 as stated. The supplementary verification in Section S1.1 claims every fixed-ν t_ν satisfies the condition; that claim conflicts with the moment bound E D_ν^r < ∞ for r<ν.","The covariance kernel's temporal factor depends only on interval fractions, not on the shape matrix, so different elliptical models are absorbed into the scalar variance and cross-ridge correlation; this suggests the null distribution of the supremum is universal within a fixed Euclidean structure, a feature worth testing for data with non-elliptical directional distributions.","For serially dependent observations, the permutation calibration used in the empirical section is no longer exact; the theory's independence assumption is the main obstacle to applying ERHT to time series with autocorrelation."],"forward_implications":["P-values from F_S are asymptotically uniform under H0, so Gaussian-supremum calibration of ERHT is valid for fixed p/n and elliptical heavy-tailed errors.","The single-change test has non-trivial local power at n^{-1/4} shifts whose spectral signal measure converges, and strong consistency when √n‖Δ‖²→∞.","The estimated break location converges at rate O_P(s_{n,ρ}^{-1}+e_{σ,n}), and if s_{n,ρ}e_{σ,n}=O(1), at rate O_P(s_{n,ρ}^{-1}).","WBS-ERHT consistently selects the number of breaks and localizes them with maximum normalized error O_P(t_WBS/(√n min_j‖Δ_j‖²)).","The ridge grid can be aggregated by the analytic Cauchy rule with exact limiting rejection probability determined by the joint Gaussian law, not by assuming the Cauchy p-value is uniform."],"fun_headline_variants":["Gaussian limit enables exact calibration for robust change-point scans","Exact calibration via Gaussian limit for heavy-tailed high-dimensional scans","Spatial medians + ridge regularization: exact break localization","New test for high-dimensional change-points: robust and exactly calibrated"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Assumption 3.1 requires the inverse radial variable ξ=R^{-1} to satisfy E ξ^{4+η}<∞ for some η>0 along with a polylogarithmic maximum bound; every uniform expansion in the proof chain depends on this moment margin, and the paper's flagship multivariate t_3 heavy-tailed design violates it because for t_ν, E ξ^r is finite only for r<ν.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian limit enables exact calibration for robust change-point scans","Exact calibration via Gaussian limit for heavy-tailed high-dimensional scans","Spatial medians + ridge regularization: exact break localization","New test for high-dimensional change-points: robust and exactly calibrated"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001448,"raw_usage":{"total_tokens":5661,"prompt_tokens":725,"completion_tokens":4936,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":4865}},"tokens_in":469,"tokens_out":4936,"duration_ms":39739,"temperature":1.0,"reasoning_tokens":4865,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T05:39:25.666506+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the ERHT single-change scan under multivariate t_3 errors with p/n≈γ and no change, compute Gaussian-supremum-calibrated p-values from F_sc, and check empirical size: if the convergence theorems were valid under t_3, rejection rates would approach α; if, as the moment gap suggests, rates deviate systematically as n,p grow, the theorems' hypotheses are not met. Equivalently, compute E(R^{-1})^{4+η} for t_3 to see it is infinite for every η>0.","supporting_citations":[],"review_version":1}