{"id":"5598559c-8956-4e64-9f47-765e65d66564","arxiv_id":"2411.14864","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Maximum pairwise Bayes factor detectors for high-dimensional Gaussian mean and covariance change points are shown to be consistent with nearly minimax localization rates.","lead":"This paper introduces Bayesian tests and estimators for change points in high-dimensional Gaussian mean vectors and covariance matrices, built from pairwise Bayes factors combined across variables and sliding windows. The methods come with consistency theorems and near-minimax localization rates, plus simulations and applications to DNA copy number and stock return data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The theorems require a fixed α satisfying (7) or α > 6C3, but the implemented method selects α by FPR calibration on data simulated from the observed sample mean/covariance; no proof connects this α̂ to the theorem conditions, so the consistency/localization guarantees do not cover the actual…","rationale":"Reading the paper in good faith, the central claim requires that the method actually used in simulations and applications is the method for which consistency and near-optimal localization are proved. The theory is built on a single-window mxPBF with a fixed prior shrinkage parameter α satisfying explicit inequalities, such as condition (7) in the mean case and α > 6C3 in the covariance case. The simulations and applications, however, use the FPR-calibrated α described in Section 4.2 and Appendix B, which is selected data-dependently by simulating N = 300 datasets from N_p(μ̂, Σ̂). The paper does not prove that this selection procedure yields an α in the theorem's range, nor that the calibration distribution approximates the true null distribution when change points are present. Without such a proof, the consistency and near-optimal localization guarantees do not apply to the implemented procedure. This is a load-bearing gap because the abstract promises guarantees for 'the proposed methods' as deployed, and the choice of α directly controls the prior shrinkage γ_nw = (nw ∨ p)^(−α), which is the mechanism behind the theorem conditions. The reader's weakest assumption identifies exactly this gap, and my independent read agrees. I do not see an additional flaw that would overturn the paper's core theoretical results for fixed α; the proofs delegate tail bounds to Lee et al. (2024), which is a reproducibility concern but not an identified error. The paper also provides reproducible code and extensive simulations, which are positive evidence, but they do not close the calibration gap. Therefore, the reader's CONDITIONAL verdict remains appropriate: the paper is worth publishing only after the FPR calibration is either proven to satisfy the theorem conditions or replaced by a provably valid choice of α.","tokens_in":31706,"tokens_out":10582,"duration_ms":113977,"concrete_test":"Run the Appendix B calibration on the Section 4.4 and 4.5 configurations (n = 500; p ∈ {200, 500, 800}; window sizes {25, 60, 100}; N = 300) over at least 100 replications, and record the selected α̂. For each configuration, check whether α̂ satisfies condition (7) for mean changes with C2 chosen as in the paper, and whether α̂ > 6C3 for covariance changes with C3 ≈ 3. Then rerun the localization evaluation with α fixed at a value inside the theoretical range (for example α = 20) and compare the F1/Hausdorff results with the FPR-calibrated results. If α̂ frequently falls outside the theorem conditions, or if the fixed-theorem-α version behaves substantially differently, the consistency guarantee does not cover the implemented algorithm.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The implemented method is not the version covered by the main theorems. Theorems 2.2 and 3.2, which support the Abstract's consistency and near-optimal localization claims, assume a fixed hyperparameter α satisfying condition (7) for the mean case or α > 6C3 for the covariance case. In contrast, Section 4.2 and Appendix B select α by an empirical false positive rate calibration: N = 300 datasets are simulated from N_p(μ̂, Σ̂), where μ̂ and Σ̂ are the sample mean and covariance of the observed sequence, and α is chosen so that the fraction of simulated mxPBFs exceeding the threshold Cth is below 0.05. No result in the paper shows that this data-dependent α̂ satisfies (7) or α̂ > 6C3, nor that calibrating on N_p(μ̂, Σ̂) transfers to the true null or alternative. The calibration is especially delicate when change points exist: the pooled sample covariance Σ̂ is generally not the covariance under the null and can be inflated by between-segment mean or covariance differences, so the simulated null distribution is not the true null distribution. Consequently, the consistency and near-optimal localization guarantees are established only for a version of the algorithm that is not the one run in the simulations or applied to the real datasets.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes new Bayesian tests and change-point estimators for high-dimensional Gaussian sequences, using maximum pairwise Bayes factors computed in local windows. For mean changes it establishes consistency of the mxPBF test (Theorem 2.1), consistent estimation of the number and locations with error nw (Theorem 2.2), and a minimax lower bound showing near-optimal localization (Theorem 2.3). Analogous results are given for covariance changes (Theorems 3.1-3.4) under conditions (A2)-(A3*). The paper also proposes a multiscale majority-voting implementation, reports simulations comparing with Dette, Geomcp, Inspect, and Edivisive, and applies the methods to CGH and S&P 500 data. The theoretical results are stated for fixed alpha satisfying condition (7) or alpha > 6C3, while the implementation selects alpha by empirical false-positive calibration.","tokens_in":32004,"tokens_out":8944,"duration_ms":90750,"significance":"If the theorems are correct, the contributions are significant: they appear to be the first Bayesian change-point methods for high-dimensional mean and covariance structures with consistency under weaker sparsity and segment-length conditions than several frequentist competitors, and with near-minimax localization rates. The paper also includes explicit minimax lower-bound constructions, a public GitHub implementation, and comparisons with four existing methods. The modularized pairwise-Bayes-factor construction is computationally attractive because it avoids p-by-p matrix inversions. However, the main theorems are not self-contained, and the implemented tuning procedure is not covered by the theoretical guarantees; these issues must be addressed before the headline claims can be accepted.","major_comments":[{"comment":"The implemented method is not the version covered by the main theorems. Theorems 2.1, 2.2, 3.1, and 3.2 require a fixed alpha satisfying condition (7) in the mean case or alpha > 6C3 in the covariance case, whereas Section 4.2 and Appendix B select alpha by empirical false-positive calibration: N=300 datasets are simulated from N_p(mu-hat, Sigma-hat), where mu-hat and Sigma-hat are the sample mean and covariance of the observed data, and alpha is chosen so that the simulated mxPBF exceeds C_th with frequency below 0.05. No result in the paper shows that this data-dependent alpha-hat satisfies condition (7) or alpha-hat > 6C3, nor that simulating from the plug-in N_p(mu-hat, Sigma-hat) is a valid null distribution when change points are present; in the covariance-change case the pooled Sigma-hat is not the null covariance. A parallel gap exists for the covariance prior constants a0, b0, b01, b02, which are fixed at 0.01 in Section 4.2 without checking the inequalities in conditions (A3) and (A3*). Consequently, the Abstract's consistency and near-optimal localization claims apply to a version of the algorithm that is not the one run in the simulations or applied to the real datasets.","section":"Section 4.2 and Appendix B; Theorems 2.1-2.2 and 3.1-3.2"},{"comment":"The key probability bounds are not proved or stated in the paper. The proofs repeatedly invoke 'the proof of Theorem 2.1 in Lee et al. (2024)' or 'the proof of Theorem 3.1 in Lee et al. (2024)' to obtain tail bounds of the form (nw or p)^{-C} for the pairwise Bayes factors. These tail bounds are exactly what powers the null and alternative consistency arguments, and they are the source of the constants C1, C2, C3, Cbm, Cbm,a and of condition (7). As written, the main theorems are not self-contained: a reader cannot verify that the quoted lemmas hold under the paper's priors, windowing setup, and high-dimensional scaling without consulting an external manuscript. The needed lemmas should be stated and proved, or at least precisely stated with assumptions sufficient for the present paper.","section":"Appendix D, proofs of Theorems 2.1 and 3.1"},{"comment":"The final bound in the proof of Theorem 3.2 says that a sum over K0 change points is at most 14 K0 (nw or p)^{-C} = o(1) for some constant 2 < C < C1. But Theorem 3.2 defines C1 as a constant arbitrarily close to, but slightly larger than, 1, so the interval (2, C1) is empty and the inequality is vacuous. The preceding step also needs an explicit relationship between C and C1,low to justify the o(1) conclusion from the assumption K0(nw or p)^{-C1,low} = o(1). This makes the proof of Theorem 3.2 invalid as written, although the theorem may be salvageable with corrected constant comparisons.","section":"Proof of Theorem 3.2, Appendix D.2"},{"comment":"The multiscale 'mxPBF major' procedure is the method used in all simulations and real-data applications, but its theoretical justification is only a heuristic sentence: if at least the window sizes n_w^(1), ..., n_w^((|N|+1)/2) satisfy the conditions of Theorems 2.2 and 3.2 and n_w^(r) <= n_w^(r+1)/2, then 'it can be shown' that the localization rate is at most n_w^((|N|+1)/2)/n. No theorem, lemma, or proof is supplied for the majority-voting interval grouping, so the finite-sample claims and the reported F1/Hausdorff performance are not connected to the proved localization rates. The inconsistency between the theoretical procedure (single window size, fixed alpha) and the evaluated procedure (multiscale majority voting, calibrated alpha) should be resolved explicitly.","section":"Section 4.1, multiscale method"}],"minor_comments":[{"comment":"The sentence 'If H_0 is true, we say that l is a change point' is backwards: under the equality null there is no change at l; the intended statement is presumably that if H_1 is true, l is a change point.","section":"Section 2.1, after equation (3)"},{"comment":"The condition '(i_k - i_{k-1}) or (i_{k+1} - i_k) >= n_w' uses a maximum where the analogous mean-structure condition in Theorem 2.1 uses a minimum; since the mxPBF at l = i_k uses n_w observations from each side, the minimum (both segment lengths at least n_w) appears to be what is needed.","section":"Theorem 3.1(ii)"},{"comment":"There is a duplicated word in 'We define Delta = Delta1 + delta I_p, where where delta = |lambda_min(Delta1)| + 0.05'; the second 'where' should be removed.","section":"Section 4.5"},{"comment":"The covariance detection method is applied after locally centering the data with a moving-average window, but Theorems 3.1-3.4 assume the data have mean exactly zero; no theory is given for the effect of this centering on the null distribution or on the localization guarantees.","section":"Section 5, real data analysis"},{"comment":"Simulation results are reported as point summaries over only 50 datasets without error bars or confidence bands; given the visible variability in F1 scores and Hausdorff distances, reporting variability or replication counts would strengthen the empirical comparisons.","section":"Figures 1-6 and Appendix A"},{"comment":"The word 'controlls' in 'which controlls the empirical false positive rate at 0.05' is a typo; it should be 'controls'.","section":"Appendix B.1"}],"recommendation":"major_revision","confidential_remarks":"The paper has a potentially strong contribution, but the central gap between the theorems (fixed alpha, single window) and the evaluated procedure (FPR-calibrated alpha, multiscale majority voting) is load-bearing and should be fixed before publication. I would also ask the authors to make the proofs self-contained with respect to the tail bounds imported from Lee et al. (2024), since those bounds drive the stated rates. The paper is within the scope of the journal; the main issue is not novelty but the degree to which the theoretical guarantees cover the actual implementation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: the paper delivers a real contribution—the first Bayesian change point detection method for high-dimensional mean and covariance that comes with consistency and near-minimax localization guarantees. The multiscale majority rule is a sensible practical fix for window choice, the simulations are extensive, and the code is on GitHub. If you work in this area, it is worth your time.\n\nThe main problem is a gap between theory and practice. Theorems 2.2 and 3.2 require a fixed alpha satisfying (7) or alpha > 6C3, but the implemented procedure selects alpha by FPR calibration on datasets simulated from the sample mean and covariance of the observed sequence, as described in Section 4.2 and Appendix B. No result shows that this data-dependent alpha satisfies the theorem conditions, and when change points exist the calibration null can be far from the true null because the pooled covariance is inflated by between-segment differences. So the consistency and localization guarantees apply to a version of the method that is not the one run in the simulations or applications. This is a load-bearing gap, not a cosmetic one.\n\nWhat the paper does well beyond the novelty: the proof strategy is internally consistent, the lower-bound arguments are standard and clean, and the comparisons, while not always apples-to-apples, are honestly reported—the authors explicitly note where direct comparison is difficult. The delegation of tail bounds to Lee et al. (2024) is acceptable but means the paper is not self-contained; a reader has to trust that prior work. The claim of \"much milder\" conditions is somewhat overstated because the signal is measured in a different norm (elementwise maximum vs. l2), so the comparison is not directly against existing results.\n\nMinor issues: the simulation figures have no error bars despite being based on 50 datasets, and the threshold Ccp is treated as a fixed default without sensitivity analysis. These are small.\n\nThis paper deserves a serious referee. The central idea is fresh, the method is implementable, and the gap is addressable: either prove the FPR-calibrated alpha satisfies the theorem conditions with high probability, or state the theorems for the data-dependent alpha directly. I would like to see this fixed before publication, but I would not desk reject it.","headline":"A genuine first Bayesian method for high-dimensional change point detection, but the implemented alpha-calibration step falls outside the theorem conditions, so the consistency claims are conditional.","tokens_in":32527,"tokens_out":2689,"would_cite":true,"duration_ms":28833,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62G20","62H15"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes the first Bayesian change point detection methods for high-dimensional mean and covariance structures, proving consistency under milder conditions and near-minimax localization rates.","keywords":["high-dimensional change point detection","Bayesian hypothesis testing","maximum pairwise Bayes factor","mean vector","covariance matrix","minimax localization rate","multiscale method","empirical false positive rate"],"falsifier":"Simulate from the null hypothesis $X_i \\sim N_p(0, I_p)$ with $n = 500$, $p = 1000$, window $n_w = 50$, threshold $C_{cp} = 10$, and $\\alpha$ chosen by the paper's FPR calibration with $N = 300$; if the proportion of null datasets in which the mxPBF exceeds $C_{cp}$ grows with $p$ instead of staying near the calibrated 0.05, then the implemented version does not have the null consistency that Theorem 2.1 claims for the theoretical $\\alpha$.","tokens_in":31484,"feed_emoji":"🎯","tokens_out":7395,"duration_ms":65595,"temperature":0.7,"pith_summary":"Change points in high-dimensional time series—shifts in the mean vector or the covariance matrix—are hard to find because the dimension can exceed the sample size and the change may involve only a few coordinates. This paper proposes the first Bayesian procedures for this problem: it tests each candidate location and each coordinate (or coordinate pair) with a windowed Bayes factor and takes the maximum as evidence of a change. The paper proves that these tests are consistent under both null and alternative, that they estimate the number and locations of change points with error bounded by the chosen window size, and that this localization rate is minimax-optimal up to a logarithmic factor. The conditions are milder than those of existing frequentist methods: the number of changed coordinates need not grow with the dimension, and no sparsity of the covariance or precision matrix is required.","feed_headline":"Bayes factors locate high-dimensional change points","feed_subtitle":"A windowed pairwise test spots mean and covariance shifts under weaker conditions, near-optimally timed.","key_machinery":"The central object is the maximum pairwise Bayes factor (mxPBF): for each candidate location $l$ and window size $n_w$, compute a Bayes factor for a change in each coordinate (mean case) or each pair of coordinates conditional on one another (covariance case) using only the $n_w$ observations before and after $l$, then take the maximum over coordinates or pairs and over candidate locations. The priors are centered at local sample means or local regression coefficients with variance scaled by $\\gamma_{n_w} = (n_w \\vee p)^{-\\alpha}$, which makes the log Bayes factors grow or shrink at polynomial rates in $n_w \\vee p$. The theory combines tail bounds on these pairwise Bayes factors with union bounds over locations and coordinates; the multiscale variant groups detections from several window sizes by a majority vote, and the localization guarantee follows because the refined estimate is chosen as the argmax within the window after the first threshold crossing.","core_discovery":"The paper establishes that change points in the mean vector or covariance matrix of a high-dimensional Gaussian sequence can be detected and localized by a maximum pairwise Bayes factor computed on small windows around each candidate location. For the mean, Theorems 2.1 and 2.2 prove that the test is consistent and that the estimated number and locations satisfy $P_0(\\hat K = K_0, \\max_k |i_k - \\hat i_k| \\le n_w) \\to 1$ whenever the minimum scaled componentwise signal exceeds roughly $\\log(n_w \\vee p)/n_w$ and certain window, sparsity, and hyperparameter conditions hold; Theorem 2.3 gives a matching minimax lower bound up to a logarithmic factor. For the covariance, Theorems 3.1 and 3.2 give analogous guarantees under signal conditions (A3) or (A3*), and Theorems 3.3 and 3.4 show that the localization rate $n_w/n$ is minimax-optimal up to $\\log(n_w \\vee p)$. The authors contend that this is the first Bayesian treatment of high-dimensional change point detection and that its conditions are weaker than those of existing frequentist approaches, in particular allowing a fixed number of changed components and requiring no sparsity of precision matrices.","pith_inferences":["The modularization into per-coordinate or per-pair Bayes factors suggests the same windowed maximum test could be applied to sub-Gaussian or heavy-tailed data after robustifying the variance estimators, but the paper's proofs rely on exact Gaussian likelihoods, so the calibration of the log Bayes factor outside normality is unverified.","Because the theory only requires one coordinate or pair with signal above the threshold, the method is intrinsically a maximum-type detector; the paper's own discussion notes it may miss many small coordinated changes, so a hybrid that first aggregates weak signals and then applies mxPBF is a natural testable extension.","The FPR-based $\\alpha$ calibration simulates from the plug-in Gaussian fit; a reader who wants to deploy the method should treat the consistency theorems as conditional on $\\alpha$ satisfying the stated inequalities and verify the calibration on their own null model.","The multiscale majority rule is proven only for window ladders with $n_w^{(r)} \\le n_w^{(r+1)}/2$ and at least half of the windows meeting the theorem conditions; in practice, the stability of detected change points across window ladders could serve as a data-driven check."],"forward_implications":["For mean changes, consistent localization within $n_w$ is guaranteed provided each change point is separated by at least $n_w$ and some coordinate moves by at least $C \\log(n_w \\vee p)/n_w$ in standardized units; the number of changed coordinates may be fixed, which existing sparse-alternative tests do not allow.","The localization error $n_w/n$ is minimax up to a logarithmic factor for single mean changes and for covariance changes under the stated signal classes, so the chosen window size directly sets the time resolution one can promise.","Covariance detection requires no sparsity of the precision matrix and no irrepresentability condition, unlike earlier tests that estimated sparse precisions.","The multiscale majority-vote version stabilizes detection across window choices, provided at least half of the window sizes satisfy the theorem conditions and the window ladder grows geometrically.","On real data, running the covariance detector first and then the mean detector within segments yields change points that align with known financial and genomic events in the two reported applications."],"supporting_citations":[{"why":"Supplies the pairwise Bayes factor derivation and the null and alternative tail probability bounds that Theorems 2.1 and 3.1 invoke directly.","marker":"Lee et al. (2024)"},{"why":"Provides the sparse-projection benchmark whose localization condition is compared in Section 2.2 and whose minimax argument Theorem 2.3 adapts.","marker":"Wang and Samworth (2018)"},{"why":"Gives the sparse-alternative change point test whose signal conditions are contrasted with condition (8) in Section 2.1.","marker":"Enikeeva and Harchaoui (2019)"},{"why":"Provides the covariance change point localization benchmark and the lower-bound construction that Theorems 3.3 and 3.4 follow.","marker":"Wang et al. (2021)"},{"why":"Is the single-change covariance estimator whose localization rate and assumptions are compared with Theorem 3.2 in Section 3.2.","marker":"Dette et al. (2022)"},{"why":"Inspires the multiscale approach and supplies the sparsity-based conditions that the covariance test in Section 3.1 improves upon.","marker":"Avanesov and Buzun (2018)"}],"fun_headline_variants":["First Bayesian high-dim change point test","Pairwise Bayes factors spot high-dim shifts","Bayesian change points with near-optimal rates","Weaker conditions for high-dim Bayesian CPD","High-dim change points via pairwise Bayes factors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theorems guarantee consistency only for hyperparameter $\\alpha$ satisfying explicit inequalities, but the implemented method selects $\\alpha$ by simulating datasets from the sample mean and covariance and picking the value whose empirical false positive rate is below 0.05; the paper does not prove that this data-driven $\\alpha$ satisfies the theoretical conditions, so the practical procedure may run outside the regime where consistency and near-optimality are proven.","fun_headline_variants_meta":{"raw":{"variants":["First Bayesian high-dim change point test","Pairwise Bayes factors spot high-dim shifts","Bayesian change points with near-optimal rates","Weaker conditions for high-dim Bayesian CPD","High-dim change points via pairwise Bayes factors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000168,"raw_usage":{"total_tokens":1253,"prompt_tokens":933,"completion_tokens":320,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":253}},"tokens_in":549,"tokens_out":320,"duration_ms":4347,"temperature":1.0,"reasoning_tokens":253,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:47:55.318216+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate from the null hypothesis $X_i \\sim N_p(0, I_p)$ with $n = 500$, $p = 1000$, window $n_w = 50$, threshold $C_{cp} = 10$, and $\\alpha$ chosen by the paper's FPR calibration with $N = 300$; if the proportion of null datasets in which the mxPBF exceeds $C_{cp}$ grows with $p$ instead of staying near the calibrated 0.05, then the implemented version does not have the null consistency that Theorem 2.1 claims for the theoretical $\\alpha$.","supporting_citations":[{"cited_title":"and Lin, L","cited_arxiv_id":null,"evidence_quote":"Supplies the pairwise Bayes factor derivation and the null and alternative tail probability bounds that Theorems 2.1 and 3.1 invoke directly."},{"cited_title":"and Harchaoui, Z","cited_arxiv_id":null,"evidence_quote":"Gives the sparse-alternative change point test whose signal conditions are contrasted with condition (8) in Section 2.1."}],"review_version":1}