{"id":"232e973c-7eb5-45eb-a80f-f2fb9a537437","arxiv_id":"2504.19306","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The authors construct spatial-median and spatial-sign CUSUM statistics for high-dimensional changepoint testing, derive their null asymptotics and asymptotic independence, and combine them adaptively via Fisher's method.","lead":"The paper builds changepoint tests for high-dimensional data that keep their statistical guarantees under heavy-tailed distributions, using spatial medians and spatial signs instead of sample means. It then combines a max-type and a sum-type statistic so the test stays powerful whether the change affects few variables strongly or many variables weakly.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (4.4) defines S† as a one-sided maximum, but Theorem 3 and its proof derive the limit of max_k |eC^T eC − p|; the factor-2 Gumbel makes the stated null distribution and p-value apply to a different statistic, explaining SSUM(0.5)'s severe under-sizing.","rationale":"The reader's weakest_assumption concerned the elliptical noise model. That is a legitimate scope limitation, but the more precise and more damaging issue is internal: the γ=0.5 max-L2 statistic is defined as a one-sided maximum, while the theorem and its proof establish a two-sided absolute-value extreme-value limit. The factor-2 Gumbel in Theorem 3 and in the p-value formula is the standard signature of a two-sided statistic. The simulation table confirms the practical consequence: SSUM(0.5) has empirical size far below nominal, which the paper attributes to slow convergence but which is exactly what one expects if a one-sided statistic is calibrated with two-sided quantiles. This does not necessarily invalidate the γ=0 results or the SCMS(0) adaptive procedure, but the central claim of the paper explicitly includes the γ=0.5 statistics, their asymptotic independence in Theorem 4(ii), and the Fisher-combined SCMS(0.5) procedure. As written, Theorem 3 and the S† p-value are not about the statistic defined in Eq. (4.4). The paper's substantial technical work on spatial-median and spatial-sign decompositions deserves credit, and the mismatch may be fixable by redefining S† as max_k |eC^T eC − p| or by deriving the correct one-sided Gumbel limit. Until that is resolved, the stated null distribution, size control, and power claims for S† and SCMS(0.5) should not be accepted as they stand.","tokens_in":56684,"tokens_out":22344,"duration_ms":254520,"concrete_test":"Run 10^4 Gaussian replications with n=200, p=200, AR(0.5) covariance, λ_n=40. Compute S† exactly from Eq. (4.4) with no absolute value inside the maximum, the two-sided statistic T† = max_k |eC_{0.5}(k)^⊤eC_{0.5}(k) − p|, and the paper's p-value using |S†| with CDF exp(−2exp(−x)). Compare the empirical rejection rate at α=0.05 for: (i) the paper formula applied to the one-sided S†; (ii) a one-sided Gumbel calibration exp(−exp(−x)) applied to the same transformed value; and (iii) the paper formula applied to T†. If (i) is roughly 0.5–1.5%, (ii) is near 5%, and (iii) is near 5%, the definition/theorem mismatch is confirmed.","verdict_should_be":"REJECT","load_bearing_attack":"The γ=0.5 max-L2 statistic is not the object whose limit is proved. Eq. (4.4) defines S†_{n,p} = max_{λ_n≤k≤n−λ_n} {eC_{0.5}(k)^⊤eC_{0.5}(k) − p}(1−n^{−1/2}), a one-sided supremum of a centered quadratic process H_{np}(k). Theorem 3 states P(A(log(n²/λ_n²))|S†|/√(2tr(R²)) ≤ x + D(...)) → exp(−2exp(−x)), and the p-value formula in Section 4 uses |S†| with the same factor-2 Gumbel. But the proof in B.3.2 replaces the statistic by max_k |H_{np}(k)|: Equations (S29)–(S31) reduce max_{1≤k≤n/2}|H_{np}(k)| to max_{λ_n≤k≤n/λ_n}|k^{−1}Z_{np}(k)|, and then Q^{(1)}_{np} takes the maximum of absolute values over two independent halves. This is the extreme-value theory for max_k |H_{np}(k)| (two-sided, factor 2), not for max_k H_{np}(k) (one-sided). The factor 2 in exp(−2exp(−x)) is exactly the signature of the absolute-value statistic. If the implemented SSUM(0.5) follows Eq. (4.4), the paper's p-value is systematically too large; Table 1 shows empirical sizes 0.2–1.6% at nominal 5%, consistent with this mismatch. Section 6.2 attributes the behavior to 'slower convergence rate', but the theorem/statistic mismatch is a more direct and verifiable explanation. Consequently, the null limit in Theorem 3, the p-value formula, and the asymptotic independence/power statements for SCMS(0.5) in Theorem 4(ii) are not established for the statistic as defined.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops robust high-dimensional changepoint tests for the mean-change model (1.1). The authors construct CUSUM-type statistics based on the spatial median (max-L-infinity tests) and on spatial signs (max-L2 tests), derive Gumbel and sup-Gaussian null limits, prove asymptotic independence between the two types of statistics, and combine the resulting p-values via Fisher's method. They also state consistency results under local alternatives and provide simulations and two real-data analyses. The central theoretical objects are the pair (M_{n,p}, S_{n,p}) and the gamma=0.5 analogue (M^dagger_{n,p}, S^dagger_{n,p}), with Theorem 4 giving the adaptive combination's null validity and Theorem 5 its local power behavior.","tokens_in":57148,"tokens_out":14072,"duration_ms":140838,"significance":"The contribution is potentially useful: it extends robust spatial-sign methodology to changepoint inference in growing dimensions, and the claimed asymptotic independence between two extreme-value limits would justify a simple Fisher combination. The paper is careful in its normalization choices and the simulations cover heavy-tailed distributions and sparsity regimes. I credit the authors for deriving explicit Gumbel normalizations and for attempting a full proof of joint convergence. However, the gamma=0.5 max-L2 branch contains a statistic/theorem mismatch that currently invalidates the stated null distribution, p-value formula, and the corresponding adaptive claims.","major_comments":[{"comment":"The statistic S^dagger_{n,p} is defined in Eq. (4.4) as a one-sided maximum, S^dagger_{n,p} = max_{\\lambda_n <= k <= n-\\lambda_n} {\\tilde C_{0.5}(k)^\\top \\tilde C_{0.5}(k) - p} (1-n^{-1/2}), without absolute values. Theorem 3 and the p-value formula in Section 4, however, use |S^dagger_{n,p}|, and the proof in §B.3.2 derives the limit of max_k |H_{np}(k)| rather than max_k H_{np}(k): Eqs. (S29)-(S31) and the definition of Q^{(1)}_{np} reduce the problem to the maximum of absolute values over two independent halves. The factor 2 in exp{-2 exp(-x)} is precisely the signature of that two-sided construction. A one-sided maximum has a different extremal index, so the null distribution in Theorem 3, the p-value formula for S^dagger_{n,p}, and the asymptotic independence/power statements in Theorems 4(ii) and 5(ii) are not established for the statistic defined in (4.4). Table 1's severe under-sizing of SSUM(0.5) (empirical sizes 0.2-1.6% at nominal 5%) is consistent with this mismatch. The explanation in Section 6.2 attributing the behavior to a 'slower convergence rate' should be revisited; the theorem/statistic mismatch is a direct and verifiable cause. The fix is either to redefine S^dagger_{n,p} with the absolute value and update the implementation, or to prove the one-sided limit and adjust the p-value and the subsequent theorems accordingly.","section":"§4, Eq. (4.4); Theorem 3; §B.3.2"},{"comment":"The simulation section sets \\lambda_n = \\lfloor 0.2 n\\rfloor for all methods (Section 6, simulation settings), but Theorems 1(ii), 2, and 3 require \\lambda_n \\sim n^\\lambda with \\lambda\\in(0,1) or \\lambda_n/n -> 0. A fixed proportion \\lambda_n = 0.2n does not satisfy these conditions. Consequently, the empirical sizes and powers reported for SMAX(0.5), SSUM(0.5), and the boundary-removal competitors are not covered by the asymptotic framework that justifies the implemented p-values. The authors should either extend the theory to fixed-proportion boundary removal, report additional simulations with \\lambda_n = n^\\lambda, or explicitly acknowledge and discuss the discrepancy.","section":"§6 simulation setup; Theorems 1-3"}],"minor_comments":[{"comment":"The phrase 'lim supp' appears to be a typo; as written, b <= limsup_p E(R_i/\\sqrt{p})^{-k} <= \\bar B only bounds the limit superior, not the sequence uniformly in p, which is what the subsequent use of \\zeta_k requires. Please state the intended uniform condition clearly.","section":"Assumption 2"},{"comment":"The p-value formula for S^dagger has an unbalanced parenthesis: D(\\log(n^2/\\lambda_n^2) should be D(\\log(n^2/\\lambda_n^2)).","section":"Section 4, p-value formula"},{"comment":"The symbol \\hat D is used both for the generic diagonal scaling in the model and for its estimator; this can confuse readers in Section 3 where \\hat D is constructed from the first and last [n\\varrho] samples. A distinct notation for the estimator would help.","section":"Section 3 and 4 notation"},{"comment":"Several secondary proofs are omitted 'by symmetry' (Theorem 5(ii), Propositions 1(ii) and 2(ii)). This is acceptable in principle, but a short sketch of the modifications needed for the omitted cases would aid verification.","section":"Theorem 5(ii), Propositions 1(ii), 2(ii)"},{"comment":"The statement that SSUM(0.5)'s under-sizing is due to the 'slower convergence rate of the statistic in SSUM(0.5)' is not supported by a rate calculation; the theorem/statistic mismatch identified in the first major comment is a more direct explanation.","section":"Section 6.2"},{"comment":"In Remark 1, the phrase 'independent with the spatial sign' should be 'independent of the spatial sign'; the grammatical slip could obscure the exact condition being assumed.","section":"Model (2.1)"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the manuscript leans heavily on a chain of self-cited results (Liu et al. 2024, Cheng et al. 2023, Feng et al. 2016) for the Bahadur representation and diagonal-matrix consistency, and the present proofs often quote these results without stating the precise versions used. The Theorem 3/statistic mismatch and the simulation-parameter mismatch are the main obstacles; both are fixable within the manuscript's scope. I do not see grounds for rejection if these points are addressed carefully."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper does something new: it builds changepoint tests on spatial medians and spatial signs, and proves that the max-L∞ statistic and the max-L2 statistic (with γ=0) are asymptotically independent, justifying Fisher's combination. That independence result is real and likely correct. Second, the γ=0.5 variant has a severe size distortion in simulations, but the stress-test explanation you flagged does not hold up: the factor-2 Gumbel in Theorem 3 comes from taking the maximum over two asymptotically independent halves, not from an absolute-value statistic. The proof works with max |H| as a technical device, but the upper tail of H is the same as the upper tail of |H| because H is bounded below; the factor 2 is from the two halves, so the theorem does apply to the stated one-sided S†. The under-sizing of SSUM(0.5) is real, but it is a finite-sample issue, not a theorem/statistic mismatch.\n\nThe paper's real contribution is the γ=0 pair: Theorem 1 (Gumbel for max-L∞), Theorem 2 (Gaussian process for max-L2), and Theorem 4(i) (asymptotic independence). These are solid and genuinely new. The proofs are long and rely heavily on self-cited results (Liu et al. 2024, Feng et al. 2016), but the dependence is transparent and the Bahadur representation is the right tool. The simulation study is extensive and the robustness gains under heavy-tailed noise are clear.\n\nSoft spots: the noise model (2.1) is elliptical-type, so the robustness claims do not cover skewed or direction-dependent heavy tails. The post-hoc exclusion of 146 stocks in the real data is a bit selective. But the biggest open question is the SSUM(0.5) size distortion: the paper's 'slower convergence' explanation is hand-wavy, and a referee should ask for a finite-sample study with larger n and a careful look at the boundary removal parameter. The asymptotic independence between two Gumbel-type limits (Theorem 4(ii)) is a nice addition, though the proof is an extension of the Gaussian case.\n\nWho this is for: anyone working on robust high-dimensional changepoint inference will want the γ=0 theory and the independence result. The γ=0.5 variant needs more empirical scrutiny before being used as a test.\n\nRecommendation: send to a serious referee. The core results deserve referee time and likely publication after revision; the size distortion must be addressed head-on.","headline":"Genuinely new independence result for spatial-median/spatial-sign changepoint tests, but the γ=0.5 variant's size distortion is real and needs a better explanation than the stress-test's mismatch claim.","tokens_in":57688,"tokens_out":15728,"would_cite":true,"duration_ms":155535,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H15","62G10","62G32","62G35"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that replacing sample means with spatial medians and spatial signs yields high-dimensional changepoint tests that are robust to heavy tails, and that combining a max-$L_\\infty$ and a max-$L_2$ statistic via Fisher's…","keywords":["changepoint detection","high-dimensional inference","spatial median","spatial sign","CUSUM statistic","asymptotic independence","Gumbel distribution","heavy-tailed robustness"],"falsifier":"Simulate $n=p=200$ under the null with zero-mean skewed heavy-tailed noise, for example independent components drawn from a centered log-normal or skew-$t$ distribution, and apply the proposed combined test at level 0.05 over many replications; if the empirical rejection rate departs markedly from 0.05 or the two normalized statistics $p^{1/2}\\hat\\zeta_1 M_{n,p}$ and $S_{n,p}/\\sqrt{2\\widehat{\\operatorname{tr}}(R^2)}$ show dependence in their joint distribution, the size and independence claims fail outside the elliptical model.","tokens_in":56486,"feed_emoji":"📊","tokens_out":9157,"duration_ms":82299,"temperature":0.7,"pith_summary":"This paper claims that high-dimensional changepoint testing can be made both robust to heavy-tailed noise and adaptive to the unknown sparsity pattern of the shift by replacing sample means with spatial medians and spatial signs. It builds two CUSUM-based statistics: a max-$L_\\infty$ statistic that detects sparse strong shifts and a max-$L_2$ statistic that detects dense weak shifts. Under the null hypothesis the first converges to a Gumbel law and the second to the supremum of a Gaussian process, and the paper shows the two limits are asymptotically independent. That independence justifies Fisher combining the two p-values, giving a test with asymptotically correct size and power at least comparable to the better single test. The practical payoff is one procedure that works across sparsity levels and heavy-tailed distributions where mean-based changepoint tests lose power.","feed_headline":"Spatial-median tests expose breaks where mean tests fail","feed_subtitle":"Combining max- and sum-type statistics keeps power for sparse and dense shifts under heavy tails.","key_machinery":"The carrying objects are the sample spatial median $\\hat\\theta_{a:b}$ with diagonal standardization, and the spatial sign $U(x)=x/\\|x\\|$ applied to standardized residuals. These feed two CUSUM families: $C_\\gamma(k)=\\{k/n(1-k/n)\\}^{1-\\gamma}\\sqrt n\\,\\hat D^{-1/2}(\\hat\\theta_{1:k}-\\hat\\theta_{k+1:n})$ for the max-$L_\\infty$ statistics $M_{n,p},M^\\dagger_{n,p}$, and a spatial-sign CUSUM $\\tilde C_\\gamma(k)$ for the max-$L_2$ statistics $S_{n,p},S^\\dagger_{n,p}$. The decisive mechanism is the Bahadur representation supplied by Liu et al. (2024), which turns each spatial median into an average of spatial signs plus a uniformly negligible remainder; that reduction lets the authors borrow the extreme-value analysis of Gaussian CUSUM processes and then transfer it to the sub-exponential spatial signs. Asymptotic independence is established by conditioning on a small block of coordinates, decomposing the remaining Gaussianized process, and bounding the interaction terms, after which Fisher's method is legitimate.","core_discovery":"The central claim is Theorem 4: under the null and Assumptions 1–3 and 6–7, the spatial-median-based max-$L_\\infty$ statistic $M_{n,p}$ and the spatial-sign-based max-$L_2$ statistic $S_{n,p}$ are asymptotically independent, with $p^{1/2}\\zeta_1 M_{n,p}$ converging to a Gumbel distribution and $S_{n,p}/\\sqrt{2\\operatorname{tr}(R^2)}$ converging weakly to the supremum of a Gaussian process $V(t)$ with covariance $\\mathbb{E}V(t)V(s)=(1-t)^2s^2$. Consequently the Fisher-combined p-value $p_{M,S}$ has asymptotically correct size and, by Theorem 5, the independence persists under a local alternative, so the combined test's power is at least comparable to the better of its two components. The same structure is proved for the boundary-removed $\\gamma=0.5$ versions $M^\\dagger_{n,p}$ and $S^\\dagger_{n,p}$, each with its own Gumbel normalization. The argument proceeds through a Bahadur representation for the sample spatial median, a Gaussian-coupling comparison of the joint law of the two statistics, and an inclusion–exclusion bound controlling their joint exceedance probabilities.","pith_inferences":["Editorial inference: the same asymptotic-independence argument should extend to any monotone combination of the two p-values, not just Fisher's, so procedures that average or take minima of transformed p-values would likely inherit the size guarantee.","Editorial inference: for skewed heavy-tailed noise (for example, log-normal or skew-$t$ arrivals), the spatial median no longer tracks the mean shift and the paper's size and power claims are not expected to hold; users should test for approximate symmetry of the noise before applying the procedure.","Editorial inference: because the spatial-sign statistic discards radius information, adding a radius-weighted term could improve power on dense alternatives; the authors' own concluding remarks point in this direction.","Editorial inference: the joint Gumbel–Gaussian-process coupling could serve as a template for other robust high-dimensional test pairs, such as rank-based or self-normalized changepoint statistics."],"forward_implications":["If the proof is right, the combined test $p_{M,S}$ has asymptotically correct size under the null and power no worse than the better of the max-$L_\\infty$ and max-$L_2$ tests, with the $\\gamma=0.5$ versions extending the guarantee when the changepoint is away from the sample edges.","The marginal limits make the tests asymptotically distribution-free: the Gumbel and Gaussian-process quantiles do not depend on the specific heavy-tailed law, only on the elliptical-type structure in Equation (2.1).","Sparse signals of order $\\|\\delta\\|_\\infty\\gtrsim\\sqrt{\\log p/n}$ are claimed detectable at the usual near-optimal rate, while dense signals are claimed detectable once $\\|\\delta\\|\\to\\infty$ under the sparsity restriction $\\|\\delta\\|^{-1}\\|\\delta\\|_\\infty=o(p^{1/2}n^{-1/2})$.","After rejection, the argmax of the chosen CUSUM statistic yields a changepoint estimator that inherits the better accuracy of the max and sum procedures across sparsity levels.","The asymptotic independence under local alternatives is the property that justifies reporting a single combined p-value rather than needing to know in advance whether the signal is sparse or dense."],"supporting_citations":[{"why":"Supplies the Bahadur representation for the sample spatial median whose remainder bound is used in every theorem's null expansion.","marker":"Liu et al. (2024)"},{"why":"The mean-based double-max-sum method whose Gumbel limits and asymptotic-independence template the paper extends to spatial statistics.","marker":"Wang and Feng (2023)"},{"why":"Provides the concentration and moment bounds for spatial signs, and the condition Assumption 1 is modeled on.","marker":"Cheng et al. (2023)"},{"why":"Introduces the scalar-transformation-invariant spatial-sign estimation with the diagonal matrix D that the CUSUM statistics rely on.","marker":"Feng et al. (2016)"},{"why":"Gives the Darling–Erdős limit for max-L2 CUSUM statistics that is the template for Theorem 3's Gumbel limit with factor 2.","marker":"Chan et al. (2013)"},{"why":"Establishes Gumbel convergence for max-L-infinity CUSUM statistics under sub-exponential noise, the template for Theorem 1.","marker":"Jirak (2015)"},{"why":"Verifies Assumption 2's radius-moment conditions for normal, t, and mixture distributions and supplies the moment framework.","marker":"Zou et al. (2014)"},{"why":"Justifies Fisher's combination of independent p-values and the power-comparability bound used for the adaptive test.","marker":"Littell and Folks (1971)"}],"fun_headline_variants":["Spatial-median tests catch breaks mean tests miss","Sparse or dense, spatial signs find change points","Adaptive changepoint test robust to heavy tails","Combine max and sum tests for stronger change detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the noise model in Equation (2.1): each noise vector must be a random radius times a common linear transformation of independent symmetric components, with the radius independent of the spatial sign; if the noise is heavy-tailed but skewed or has direction-dependent scale, the spatial median may not estimate the mean shift and the Gumbel, Gaussian-process, and independence limits can fail.","fun_headline_variants_meta":{"raw":{"variants":["Spatial-median tests catch breaks mean tests miss","Sparse or dense, spatial signs find change points","Adaptive changepoint test robust to heavy tails","Combine max and sum tests for stronger change detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000694,"raw_usage":{"total_tokens":3161,"prompt_tokens":987,"completion_tokens":2174,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":2111}},"tokens_in":603,"tokens_out":2174,"duration_ms":16818,"temperature":1.0,"reasoning_tokens":2111,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:55:29.280223+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate $n=p=200$ under the null with zero-mean skewed heavy-tailed noise, for example independent components drawn from a centered log-normal or skew-$t$ distribution, and apply the proposed combined test at level 0.05 over many replications; if the empirical rejection rate departs markedly from 0.05 or the two normalized statistics $p^{1/2}\\hat\\zeta_1 M_{n,p}$ and $S_{n,p}/\\sqrt{2\\widehat{\\operatorname{tr}}(R^2)}$ show dependence in their joint distribution, the size and independence claims fail outside the elliptical model.","supporting_citations":[{"cited_title":"and Feng, L","cited_arxiv_id":null,"evidence_quote":"The mean-based double-max-sum method whose Gumbel limits and asymptotic-independence template the paper extends to spatial statistics."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the Darling–Erdős limit for max-L2 CUSUM statistics that is the template for Theorem 3's Gumbel limit with factor 2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes Gumbel convergence for max-L-infinity CUSUM statistics under sub-exponential noise, the template for Theorem 1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Justifies Fisher's combination of independent p-values and the power-comparability bound used for the adaptive test."}],"review_version":1}