{"id":"f4ba2c4c-7860-40e0-b678-56175c776f78","arxiv_id":"2607.16106","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"AR(p)-FOCuS is an exact online generalised-likelihood-ratio detector for mean changes under AR(p) noise, with an average cost close to O(log n) per new observation.","lead":"This paper extends the FOCuS online changepoint detection algorithm to data with autocorrelated noise modelled by an AR(p) process. The new AR(p)-FOCuS method is designed to detect mean shifts in streaming data quickly and at low computational cost, with applications to network traffic monitoring and similar high-frequency data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The power advantage of AR(p)-FOCuS is established only under a correctly specified AR(p) null; the paper's own misspecification results show the advantage can reverse, so the abstract's broad 'autocorrelated data' claim needs qualification.","rationale":"The mathematical derivation of Proposition 1 is internally consistent and the exactness of the GLR statistic under known AR(p) parameters is credible; the simulations under AR(p) noise support the power gain. The weakest point is the transfer of this exactness to the claimed practical advantage: parameter estimation from a short probation period and the possibility of non-AR structure in real data. The paper is honest about some of this (Figures 4 and 16, Discussion), but the abstract states a broad claim. The reader's conditional verdict captures this; I agree and propose a targeted simulation under non-AR alternatives as the decisive check. This is a scope/robustness concern, not a detected error in the algorithm, so no verdict change is needed.","tokens_in":23036,"tokens_out":20374,"duration_ms":181460,"concrete_test":"Simulate null and alternative data from a non-AR autocorrelated process, e.g. ARMA(1,1) with phi=0.9, theta=0.5, and also AR(1) plus a linear trend. For each, use the Section 3.2 Monte Carlo procedure to calibrate thresholds for AR(p)-FOCuS, IID FOCuS, and FOCuS-prewhiten at the same false-positive rate under the true null. Compare detection delays and, separately, the empirical false-positive rate of AR(p)-FOCuS when the threshold was calibrated under a misspecified fitted AR null. If AR(p)-FOCuS no longer dominates, or its FPR exceeds the nominal level, the central advantage requires explicit restriction to AR(p) noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that AR(p)-FOCuS has greater detection power than IID tests when data are autocorrelated—depends on the noise being exactly AR(p) with order and coefficients well estimated from a change-free probation period. Section 3.2 estimates these parameters, and the simulation study in Section 4.2 (Figure 4) shows that with short probation AR(p)-FOCuS can be beaten by plain FOCuS; Appendix E.3 (Figure 16) shows that imposing an AR(1) coefficient on IID data destroys power for small changes. The real-data evaluation in Section 5 has no ground truth: thresholds are calibrated under the fitted AR null, and Table 1 compares only commonly detected changes using min of the two estimated changepoints as the reference. If the telemetry noise is not AR(p) (the introduction itself mentions slow trends and non-AR dependence), the higher detection counts may reflect threshold miscalibration rather than sensitivity. This is a scope limitation rather than an internal inconsistency, but it makes the abstract's broad 'autocorrelation' claim the least secure part of the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends the FOCuS online changepoint-detection algorithm to observations whose noise follows an AR(p) process. The main methodological content is a recursion (Proposition 1) that expresses the generalized likelihood-ratio statistic for a change in mean under AR(p) noise as a maximum of a small set of quadratic functions, thus permitting functional pruning. An extension to unknown pre-change mean and a pruning rule for that case (Proposition 2) are also given, together with a Monte Carlo calibration procedure that accounts for AR parameter estimation from a probation period. Simulation studies compare the proposed AR(p)-FOCuS with IID FOCuS and with prewhitening-based FOCuS under various AR processes and probation lengths, and a real telecommunications dataset is analysed. The paper claims an average O(log n) per-iteration cost, based on the observed number of stored quadratics, and greater detection power than IID-based tests on autocorrelated data.","tokens_in":23380,"tokens_out":6532,"duration_ms":57241,"significance":"If the central claims hold, the paper is a practically useful extension of FOCuS: it gives a clean derivation of the GLR statistic for AR(p) noise, a functional pruning recursion that appears correct, an explicit treatment of unknown pre-change mean, and a threshold calibration procedure that propagates estimation uncertainty from the probation period. The inclusion of replication code and the honest discussion of estimation sensitivity (Sections 4.2, 6, and Appendix E.3) are strengths. However, the two headline claims—average O(log n) complexity and power gains on autocorrelated data—are currently supported only partially: the complexity result is empirical, and the power gains are conditional on correct or well-estimated AR(p) structure. The real-data application supplies no ground truth, so the reported detection-count improvements are not yet conclusive evidence of sensitivity.","major_comments":[{"comment":"","section":"§4.1, Figure 3, and abstract"},{"comment":"","section":"§4.2 and Appendix E.3, Figure 16"},{"comment":"","section":"§5, Table 1"}],"minor_comments":[{"comment":"","section":"Appendix B"},{"comment":"","section":"References and text"},{"comment":"","section":"Figure 3 caption and §4.1"},{"comment":"","section":"§3.1"},{"comment":"","section":"Appendix E.1, Figure 13 caption"}],"recommendation":"major_revision","confidential_remarks":"The core derivation (Proposition 1) appears sound and the paper is a credible extension of FOCuS. The main obstacles are the unproved complexity claim, the overbroad wording of the power claim in the abstract, and the lack of ground truth in the real-data section. These are fixable by adding a theorem or weakening the claims, and by adding explicit caveats; hence major revision rather than rejection. I would also recommend that the authors state clearly which parts of the complexity claim are conjectural, since the paper's contribution is partly computational."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First: this is a genuine, useful extension of FOCuS to AR(p) noise. The recursion separating the last p observations from older changes (Proposition 1) is the natural way to handle the transient and, as far as I can tell, correct. The authors also do the right thing by acknowledging that the GLR statistic itself goes back to Lai and Gombay; the contribution is the pruning-based implementation that makes it feasible online. The added value is real.\n\nThe paper's strengths: the derivation is clean, the Appendix proofs are adequate, the simulation design is careful (known vs estimated AR parameters, order selection by AIC, probation length, and a misspecification experiment in the supplement). The Monte Carlo threshold calibration that re-estimates parameters from simulated probation periods is a nice touch. They ship R code. That is reproducible evidence and should count in its favor.\n\nNow the soft spots. The headline claim of average O(log n) per-iteration cost is supported only by simulation (Figure 3). The proof for AR(p) is not given; the authors borrow the pruning argument from the IID FOCuS paper without showing the expected number of candidate functions remains O(log n) under dependence. This may be true, but as stated it is an empirical observation, not a theorem.\n\nThe second issue is scope. The abstract says 'greater detection power' for autocorrelated data. That is true only when the noise is genuinely AR(p) and the parameters are well estimated. The paper's own Figure 4 shows that with a short probation period AR(p)-FOCuS is beaten by plain FOCuS, and Figure 16 shows that imposing an AR coefficient on IID data kills power for small changes. Real streams often have slow trends or non-AR dependence, as the introduction itself notes. So the power advantage is conditional, and the abstract should say so.\n\nThird, the real-data section has no ground truth. Table 1 compares only commonly detected changes and uses the earlier of the two estimated changepoints as the reference. Thresholds are calibrated under the fitted AR null; if the telemetry is not AR(p), the higher detection counts could reflect calibration error rather than sensitivity. This is a limitation, not a fatal flaw, but it deserves a caveat in the text.\n\nOverall, this is a solid computational statistics paper for anyone working on online changepoint detection with autocorrelated streams. The core algorithm is sound, the experiments are extensive, and the authors are honest about limitations. It deserves a serious referee. A good referee should press for a proof or a carefully qualified statement of the complexity claim, and a more measured abstract. I'd cite it for the recursion and the algorithm.","headline":"Solid algorithmic extension of FOCuS to AR(p) noise; the O(log n) claim and the power claim are less secure than the abstract suggests.","tokens_in":23799,"tokens_out":3529,"would_cite":true,"duration_ms":31516,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62L10","62M10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper presents AR(p)-FOCuS, an exact and computationally efficient implementation of the generalized likelihood-ratio test for online detection of mean changes when the noise is an autoregressive process of order p, achieving average O","keywords":["autoregressive processes","online changepoint detection","generalized likelihood-ratio test","functional pruning","CUSUM","autocorrelation","streaming data","sequential detection"],"falsifier":"Simulate data that is not AR(p)—for instance, long-range dependence or a slow linear trend—then calibrate the threshold using the paper's constant-mean AR(p) null and check whether the empirical false-positive rate stays at the nominal level. Alternatively, feed truly IID data while assuming a strong AR(1) coefficient (e.g., ρ=0.9) and observe detection power for small changes: the paper's Figure 16 indicates it can fall to near zero, which would directly expose the sensitivity to misspecification.","tokens_in":22976,"feed_emoji":"📈","tokens_out":3697,"duration_ms":35874,"temperature":0.7,"pith_summary":"The paper claims that the generalized likelihood-ratio (GLR) test for a change in mean can be computed exactly and efficiently when the noise follows an autoregressive process of order p, by adapting the functional-pruning technique behind the FOCuS algorithm. The key recursion isolates changes within the last p time steps from older changes, allowing candidate changepoints that can never be optimal to be pruned. The resulting AR(p)-FOCuS algorithm has average O(log n) computational cost per iteration and, in simulations, detects changes faster than IID-based methods when the data are autocorrelated. This matters for real-time monitoring of high-frequency data streams—such as network telemetry—where serial dependence is common and fast, exact detection is needed.","feed_headline":"Exact online changepoint detection for autocorrelated streams","feed_subtitle":"New recursion isolates recent and old changes, pruning candidates so detection stays fast and exact even under AR(p) noise.","key_machinery":"The key machinery is the recursion of Proposition 1, which splits the likelihood-ratio statistic into two parts: the recent-change terms S_{n,n−j}(μ1) for j=1,…,p, which capture the transient effect of a mean shift on the first p whitened observations, and the older-change term Q_n(μ1), which is updated recursively. This split allows the FOCuS-style functional pruning to operate only on Q_n, where the post-change mean has reached its long-run level. The algorithm stores the statistic as a set of quadratics in μ1, each corresponding to a candidate changepoint, and prunes those that cannot be optimal for any future data. The pruning is the mechanism that keeps the per-iteration cost at O(log n","core_discovery":"AR(p)-FOCuS is an exact implementation of the GLR test for a change in mean when the noise is AR(p), not an approximation. The central identity is Proposition 1: LR_n(μ1) = max( max_{j=1..p} S_{n,n−j}(μ1), Q_n(μ1) ), where Q_n satisfies the recursion Q_n(μ1) = max{ S_{n−1,n−p−1}(μ1), Q_{n−1}(μ1) } + C_n(μ1). This separates the likelihood-ratio statistic into contributions from changes that occurred within the last p observations—handled directly—and older changes, which are propagated recursively. The recursion enables functional pruning: the algorithm stores the statistic as a piecewise-quadratic function of the post-change mean and discards quadratics that can never become maximal, so the","pith_inferences":["Inference: The recursion's separation of 'recent' and 'old' changes is a structural idea that could extend to other stationary noise models (e.g., ARMA) where the likelihood ratio has a similar finite-memory transient.","Inference: Because the detection threshold is calibrated under a constant-mean AR(p) null, applying the method to streams with slow trends or non-AR dependence may inflate false positives; a model diagnostic or a more robust calibration (e.g., block bootstrap) would be a prudent extension.","Inference: The logarithmic storage guarantee likely depends on the quadratic form of the Gaussian likelihood ratio; testing sub-Gaussian or heavy-tailed noise would reveal whether the pruning property is robust beyond the Gaussian case."],"forward_implications":["AR(p)-FOCuS provides an exact online GLR test for mean changes under AR(p) noise, so practitioners no longer need to ignore autocorrelation or rely on approximate pre-whitening.","The average O(log n) per-iteration cost makes the method practical for high-frequency data streams where a naive GLR computation would be infeasible.","Simulations show that modelling the AR structure reduces detection delay substantially when autocorrelation is strong, while remaining competitive with IID-based methods when autocorrelation is weak.","When AR parameters are estimated from a probation period, performance converges to the oracle version as the probation period grows, and over-specifying the AR order is safer than under-specifying it.","On the telecommunications datasets, AR(p)-FOCuS detects more changepoints and achieves lower average delays than both IID FOCuS and pre-whitened FOCuS on commonly detected changes."],"fun_headline_variants":["Exact GLR changepoint detection for autocorrelated streams","AR(p) changepoint detection: exact GLR, O(log n) per step","Pruned recursion brings exact online changepoint detection to AR(p)","Online changepoint detection that stays exact under autocorrelation","Watch for shifts in autocorrelated data with exact GLR"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method's advantage rests on the assumption that the true noise is genuinely AR(p) and that its order and coefficients can be estimated reliably from a change-free probation period; if that period is short or the model is misspecified, the paper's own simulations show it can perform worse than the plain IID detector.","fun_headline_variants_meta":{"raw":{"variants":["Exact GLR changepoint detection for autocorrelated streams","AR(p) changepoint detection: exact GLR, O(log n) per step","Pruned recursion brings exact online changepoint detection to AR(p)","Online changepoint detection that stays exact under autocorrelation","Watch for shifts in autocorrelated data with exact GLR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001015,"raw_usage":{"total_tokens":4128,"prompt_tokens":754,"completion_tokens":3374,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":3294}},"tokens_in":498,"tokens_out":3374,"duration_ms":21796,"temperature":1.0,"reasoning_tokens":3294,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T21:18:46.422822+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate data that is not AR(p)—for instance, long-range dependence or a slow linear trend—then calibrate the threshold using the paper's constant-mean AR(p) null and check whether the empirical false-positive rate stays at the nominal level. Alternatively, feed truly IID data while assuming a strong AR(1) coefficient (e.g., ρ=0.9) and observe detection power for small changes: the paper's Figure 16 indicates it can fall to near zero, which would directly expose the sensitivity to misspecification.","supporting_citations":[],"review_version":1}