{"id":"c256b835-5c67-4756-88bc-c83a0c2e8c1b","arxiv_id":"2608.13352","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"BARBS is a bootstrap-calibrated binary segmentation method that detects multiple change points in nonstationary dependent time series with Type I error control and near-optimal localization rates.","lead":"This paper introduces BARBS, a recursive binary segmentation method that detects multiple change points in time series even when the noise is nonstationary and dependent. It uses a bootstrap to calibrate test thresholds, controls false positives, and provides confidence intervals for change point locations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Section 5.1 plug-in rule for m is not shown to satisfy Assumption (A4), so Theorem 3.1's Type I and localization guarantees are unproven for BARBS as implemented.","rationale":"The reader's weakest_assumption already targets the plug-in m, and my re-reading agrees. The strongest claim is Theorem 3.1(b); its proof (supplement) is not available, but the more specific and falsifiable weakness is the mismatch between the theorem's conditions and the algorithm's tuning. A theorem conditional on an unverifiable, unestimated condition cannot underwrite the practical claim. The concern is not that the asymptotic theory is wrong, but that it has not been connected to the implemented procedure. Other possible concerns—the asymptotic-independence heuristic in Section 2.3, the fixed B=2000 quantile resolution—are real limitations but secondary: the alpha-spending upper bound still gives false-positive control, and B can in principle be increased with n. The m gap is primary because it invalidates the precondition of the theorem. The proposed computational check is targeted and would settle whether the gap manifests in finite samples. For these reasons I recommend no change to the reader's CONDITIONAL verdict.","tokens_in":17394,"tokens_out":18092,"duration_ms":195014,"concrete_test":"Use a DGP satisfying Assumptions (A1)-(A2) with q=4, ϖ=0, and Θ=0.95, so Assumption (A4) requires 0.5<ω<0.6. Generate n=10^4 and n=10^5 with change points spaced n^Θ apart. Compute the Section 5.1 plug-in m on each replication and record log m/log n; check whether it is in (0.5,0.6) and does not converge to 1/3. Then run BARBS with the plug-in m and, in parallel, with m=⌊n^{0.55}⌋, using B=5000 to avoid a quantile-estimation confound, and compare the empirical frequency of the Theorem 3.1(b) event {brn=rn and max_l |bkl−k∗l|≤Cγ_n^{-2}Δ_n^{-2} log n} over 500 replications at α=0.05. If the plug-in m gives materially lower frequency than the admissible m, the concern is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing condition for Theorem 3.1(b) is Assumption (A4), which requires the window size m≍n^ω to lie in a specific open interval determined by the moment q, the minimum spacing exponent Θ, and the jump exponent ϖ. The theorem is conditional on this choice. The implemented BARBS, however, selects m by the plug-in rule in Section 5.1, which is designed to minimize the MSE of a long-run variance estimator and targets m≍n^{1/3}. No lemma or remark establishes that this data-driven m falls in the admissible interval with high probability, or that the asymptotic results hold uniformly over data-dependent m. This is not a merely formal gap: for q=4, ϖ=0, Θ=0.95, (A4) requires 0.5<ω<0.6, so the plug-in target ω=1/3 is outside the admissible range. Even for large q, the admissible band is nonempty only for Θ sufficiently far from the boundary 6/7+4ϖ/7, and the plug-in rule does not adapt to Θ or ϖ. Hence the central claim—simultaneous false-positive control and recovery with the stated localization rate—is not supported for the procedure as actually run in Sections 5 and 6. A second practical gap (Remark 3: fixed B=2000 cannot estimate the smallest tail probabilities when α is fixed and γn→0) reinforces the diagnosis, but the m gap is primary because it directly breaks the assumption stated in the theorem.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes bootstrap-assisted robust binary segmentation (BARBS), a recursive binary segmentation method for detecting multiple change points in time series with mean changes under nonstationary and weakly dependent noise. The method computes CUSUM statistics on candidate intervals, calibrates critical values using a Gaussian multiplier bootstrap, and allocates the overall significance level α across intervals in proportion to interval length. The main theoretical results are: (a) under the null, the probability of declaring at least one change point is asymptotically α; (b) under an alternative with minimum jump magnitude Δ_n and minimum spacing γ_n, BARBS recovers the true number of change points and localizes each estimated change point to within O(γ_n^{-2} Δ_n^{-2} log n) with probability at least 1 − α(1+o(1)); (c) a second-stage refinement produces estimators with limiting distributions (Theorems 4.1 and 4.2) and a nearly optimal uniform localization rate (Theorem 4.3). The paper includes an extensive simulation study covering seven nonstationary data-generating processes and an application to U.S. inflation data.","tokens_in":17682,"tokens_out":12215,"duration_ms":104255,"significance":"If the results are correct, the paper makes a substantial contribution by extending multiple change point detection to genuinely nonstationary error processes with finite moments and physical dependence, going beyond the locally stationary or piecewise locally stationary assumptions common in this literature. The length-weighted Bonferroni allocation is a natural way to account for the multiple-testing structure in recursive segmentation, and the bootstrap calibration is a useful practical device. The simulation study is thorough, with multiple scenarios, several competing methods, and metrics that distinguish detection power from false-positive control. The main weakness is that the theoretical guarantees are conditional on Assumption (A4), but the implemented tuning parameter choice is not shown to satisfy that assumption, so the theorems do not cover the procedure as actually run.","major_comments":[{"comment":"The plug-in rule selects m = [ (2 \\hat{C}_{2n} n / \\hat{C}_{1n})^{1/3} ], which is of the order n^{1/3}. Assumption (A4) requires m \\asymp n^\\omega with \\omega in an interval whose lower endpoint is max(2/q, 3−3Θ, 2−2Θ+2ϖ, 9−10Θ+6ϖ) and upper endpoint (7+4/q)Θ − (6+4/q). For q=4, ϖ=0, Θ=0.95, this interval is (0.5, 0.6), so ω=1/3 is outside. The paper contains no lemma showing that the data-driven m satisfies (A4) with high probability, nor an argument that Theorem 3.1 holds uniformly over data-dependent m. Because Theorem 3.1(a)-(b) are stated under (A1)-(A4), the Type I error control and recovery guarantees are not established for BARBS as implemented in Sections 5 and 6. This is a load-bearing gap: the authors should either modify the procedure so that the selected m provably lies in the admissible range, or state conditions on the data-generating process under which the plug-in rule is admissible and restrict the claims accordingly.","section":"Section 5.1 and Assumption (A4)"},{"comment":"The theoretical statements allow α = α_n → 0, and Remark 3 itself notes that stable estimation of the extreme quantile requires B α_n γ_n → ∞. However, Section 5.1 fixes B = 2000, so for any α_n → 0 this condition eventually fails. The paper does not state how B must depend on n in Theorems 3.1 and 4.1-4.3, nor whether the proofs treat B as part of the procedure. If the intended scope is fixed α (as in the simulations), the statements and remarks should say so explicitly; if α_n → 0 is claimed, the bootstrap replication number should be included in the procedure and its growth rate specified.","section":"Remark 3 and Section 5.1"},{"comment":"The proofs of Theorems 3.1 and 4.1-4.3 are all contained in a supplementary file that was not available to the referee. In particular, the argument behind the 'α{1+O(1)}' statement in Theorem 3.1 requires asymptotic independence of the terminal tests on disjoint intervals, which is plausible but needs a careful proof. The referee could not verify the correctness of the central results; the supplement should be made available with the revision.","section":"Theorems 3.1, 4.1-4.3 (supplementary proofs)"}],"minor_comments":[{"comment":"The quantity m is called 'block size' in the text of Section 5.1 but 'window size' elsewhere; please use one consistent term.","section":"Section 5.1"},{"comment":"The confidence interval coverage results in Table 5 are conditional on successful detection of the change points; this conditioning should be stated in the table caption.","section":"Section 5.6"},{"comment":"The values of α and m used for the real data analysis are not reported; both are user-specified tuning parameters and should be given for reproducibility.","section":"Section 6"},{"comment":"In the description of the U.S. inflation data, 'U.S. City A verage' contains an unintended space; also 'efficiency' and other ligature artifacts appear in a few places and should be cleaned up.","section":"Section 6 and throughout"},{"comment":"The bootstrap statistic S^{(b)}_{s,k,e} uses a different normalizing sequence than T_{s:e,m}; a short remark justifying this choice would help the reader.","section":"Section 2.2"},{"comment":"The phrase 'our condition in Assumption (A4) requires at least Θ>9/10' is only true in the limit q→∞; for finite q, the exact threshold depends on q and ϖ. Please clarify the statement.","section":"Section 3, paragraph after Theorem 3.1"}],"recommendation":"major_revision","confidential_remarks":"The central gap—the mismatch between the theoretically required window size order and the plug-in rule—is the main obstacle. I believe the paper is salvageable and the ideas are promising; I would be willing to re-review a revision that closes this gap, either by a data-driven m selection that provably satisfies (A4) or by restricting the theoretical claims to the range where n^{1/3} is admissible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a real contribution to multiple change point detection under nonstationary noise. BARBS—binary segmentation with Gaussian multiplier bootstrap critical values and a length-weighted Bonferroni allocation—is new, clearly described, and honestly tested. The simulation study is extensive: seven noise models, several signal configurations, Type I error under the null, and confidence interval coverage, with the known weakness in the PLS scenario openly reported. The theory is precise, with a clear set of assumptions. That said, there is a real gap between the theory and the implementation. Theorem 3.1(b) requires the window size m to satisfy Assumption (A4), which for q=4 and Θ=0.95 forces m ~ n^ω with ω in (0.5, 0.6). The plug-in rule in Section 5.1 targets m ~ n^{1/3}, which is outside that range. No result in the paper shows that the data-driven m satisfies (A4) with high probability, or that the theorem is uniform over data-dependent m. So the main guarantee—false positive control plus recovery of all change points with the stated localization rate—is not actually proved for the procedure as run in the simulations and the data application. The paper also relies on a heuristic 'asymptotic independence' of terminal tests in Section 2.3; the Bonferroni budget alone may suffice, but the claim is not checked. These are not fatal objections to the method; the simulations suggest the procedure works well in practice. But the theory and implementation need to be reconciled before the headline claims are accepted. The supplement was not available to me, so I cannot comment on the proofs. Who is this for: statisticians and econometricians working on change points under complex temporal dependence. The paper deserves a serious referee: the idea is good, the experiments are careful, and the gap is fixable, either by proving the plug-in m works under additional conditions, or by adjusting the implementation to match the assumptions. My recommendation: engage with it, but require the authors to close the m gap.","headline":"A genuinely useful method with an honest simulation study, but the main theorem's window-size condition is not met by the plug-in rule used in practice, so the headline guarantee is unproven for the implemented procedure.","tokens_in":18192,"tokens_out":3479,"would_cite":false,"duration_ms":31481,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M10","62G09","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A bootstrap-assisted recursive scan detects multiple mean breaks in nonstationary time series while controlling the global false-positive rate.","keywords":["change point detection","binary segmentation","Gaussian multiplier bootstrap","CUSUM","physical dependence","nonstationary time series","confidence intervals","Type I error control"],"falsifier":"Run BARBS at $\\alpha=0.05$ on $n=5000$ observations of nonstationary noise with no mean break, using a window size $m$ just outside the admissible range of Assumption (A4), for instance $m$ of order $n^{0.9}$ when the spacing condition forces $\\omega<0.9$, and check whether the empirical rejection rate remains near $0.05$ rather than drifting upward.","tokens_in":17170,"feed_emoji":"📈","tokens_out":9932,"duration_ms":80910,"temperature":0.7,"pith_summary":"The paper introduces BARBS, a recursive binary-segmentation procedure for detecting multiple changes in the mean of a time series when the noise is allowed to be nonstationary and dependent. The central claim is that BARBS asymptotically controls the overall false positive probability at any prespecified level $\\alpha$ when no changes are present, and that when changes exist it recovers the correct number with probability at least $1-\\alpha$ while localizing every estimated break within a stated error bound. A second-stage refinement is shown to achieve the optimal individual localization rate for each break and to provide asymptotic distributions from which confidence intervals can be constructed. This matters because existing multiple change point methods are largely calibrated under i.i.d. or stationary noise, which yields inflated false positives in nonstationary series.","feed_headline":"Bootstrap scan controls false positives in change point detection","feed_subtitle":"New binary-segmentation procedure locates multiple mean shifts in nonstationary time series with calibrated error rates.","key_machinery":"The central object is the Gaussian multiplier bootstrap applied to weighted CUSUM statistics on each interval. For an interval $[s,e]$, the statistic $T_{s:e,m}$ scans a CUSUM process over candidate breakpoints, and the bootstrap statistic $T^{(b)}_{s:e,m}$ is built from local moving-window differences $Y_j$ multiplied by i.i.d. standard Gaussians; its conditional quantile supplies the critical value. The procedure spreads the global error budget $\\alpha$ across intervals in proportion to interval length, so each terminal interval carries local level $(e-s+1)\\alpha/n$, a length-weighted Bonferroni correction. This calibration machinery is what lets the recursion stop at the right place: it avoids the uncalibrated thresholds of earlier segmentation methods and avoids false-positive inflation under nonstationary dependence.","core_discovery":"On the paper's own terms, the discovery is that a carefully budgeted sequence of Gaussian multiplier bootstrap CUSUM tests can carry out binary segmentation with valid error control under genuinely nonstationary temporal dynamics. In the absence of change points, Theorem 3.1(a) states $\\Pr(\\hat r_n>0)=\\alpha\\{1+O(1)\\}$. When $r_n\\ge1$ true breaks exist with minimum spacing $\\gamma_n$ and minimum jump $\\Delta_n$, Theorem 3.1(b) states that with probability at least $1-\\alpha\\{1+O(1)\\}$ BARBS returns exactly $r_n$ breaks and every estimate satisfies $|\\hat k_l-k_l^*|\\le C\\gamma_n^{-2}\\Delta_n^{-2}\\log n$. The refined estimators satisfy the distributional limits of Theorems 4.1 and 4.2 and the near-optimal uniform rate of Theorem 4.3. Together these results assert that a practitioner can specify a global significance level and obtain both detection power and honest uncertainty quantification in settings where previous binary-segmentation theory required stationarity or i.i.d. errors.","pith_inferences":["A testable extension would be to transplant the length-proportional significance allocation into seeded binary segmentation or narrowest-over-threshold search, potentially giving those algorithms a calibrated stopping rule to replace threshold-plus-model-selection heuristics.","The data-driven plug-in rule for the window size $m$ is not proven to land inside the Assumption (A4) region; a small simulation experiment varying $m$ around the plug-in value could reveal how sensitive the nominal Type I error guarantee is.","Because the Type I error bound relies on asymptotic independence of tests on disjoint terminal intervals, long-range dependence or strong periodic dependence could make the $O(1)$ correction in Theorem 3.1 nontrivial; checking BARBS under fractionally integrated noise would delimit the guarantee.","For shrinking jumps, Theorem 4.2 requires piecewise local stationarity with long-run variance continuous at the break, and the paper's own simulation shows coverage drops when a break coincides with a variance discontinuity; confidence intervals should be reported with that caveat in mind."],"forward_implications":["A user can run BARBS at a chosen $\\alpha$ and expect the probability of any spurious break to be asymptotically $\\alpha$, even when the noise is locally stationary, piecewise stationary, autoregressive, or nonlinear.","When the minimum spacing and jump size meet the Assumption (A4) window, the procedure simultaneously finds the true number of breaks and a uniform localization error of order $\\gamma_n^{-2}\\Delta_n^{-2}\\log n$.","The refined estimators reach the optimal individual localization rate in fixed-jump cases and a near-optimal rate when jumps shrink, and they admit confidence intervals based on the limit distributions in Theorem 4.2.","Implementation is guided by the requirement that the number of bootstrap replications satisfy $B\\alpha\\gamma_n\\to\\infty$ before empirical quantiles at extreme levels are trusted."],"supporting_citations":[{"why":"Defines the CUSUM statistic for heteroscedasticity and autocorrelation robust structural change detection that BARBS adapts to every interval.","marker":"Zhou (2013)"},{"why":"Introduces the physical dependence measure used in Assumption (A1) to quantify nonstationary temporal dependence.","marker":"Wu (2005)"},{"why":"Supplies probability and moment inequalities under dependence used in proving the bootstrap calibration.","marker":"Liu et al. (2013)"},{"why":"Provides the Gaussian comparison and anti-concentration results and the plug-in rule for selecting the window size $m$.","marker":"Wu and Zhou (2024a)"},{"why":"Supplies the binary segmentation framework, the second-stage refinement estimator (4.1), and the benchmark test that BARBS overcomes under nonstationarity.","marker":"Bai (1997)"},{"why":"Provides the wild binary segmentation benchmark and the uniform convergence rate counterpart under i.i.d. Gaussian errors.","marker":"Fryzlewicz (2014)"},{"why":"Establishes the optimal localization rate benchmark and multiscale jump testing under complex temporal dynamics used to assess Theorem 4.3.","marker":"Wu and Zhou (2024b)"},{"why":"Provides the Gaussian approximation for nonstationary time series underlying the sequential approximation in Lemma A.1.","marker":"Bonnerjee et al. (2024)"}],"fun_headline_variants":["BARBS: bootstrap binary segmentation with calibrated error control","Nonstationary change point detection with robust bootstrap tests","Error-controlled multiple break detection via recursive CUSUM","Optimal localization for multiple abrupt shifts in time series","Bootstrap-assisted segmentation finds true breaks in dependent data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theory requires the window size $m$ to fall inside a narrow asymptotic range set by the minimum spacing between breaks; the implemented plug-in rule for $m$ is not shown to land in that range, so if it does not, the false-positive and recovery guarantees could fail in practice.","fun_headline_variants_meta":{"raw":{"variants":["BARBS: bootstrap binary segmentation with calibrated error control","Nonstationary change point detection with robust bootstrap tests","Error-controlled multiple break detection via recursive CUSUM","Optimal localization for multiple abrupt shifts in time series","Bootstrap-assisted segmentation finds true breaks in dependent data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001008,"raw_usage":{"total_tokens":4270,"prompt_tokens":961,"completion_tokens":3309,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":3233}},"tokens_in":577,"tokens_out":3309,"duration_ms":18651,"temperature":1.0,"reasoning_tokens":3233,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:05:32.081259+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run BARBS at $\\alpha=0.05$ on $n=5000$ observations of nonstationary noise with no mean break, using a window size $m$ just outside the admissible range of Assumption (A4), for instance $m$ of order $n^{0.9}$ when the spacing condition forces $\\omega<0.9$, and check whether the empirical rejection rate remains near $0.05$ rather than drifting upward.","supporting_citations":[],"review_version":1}