{"id":"56a82e1b-b8f1-46c1-a155-b32cd45c570c","arxiv_id":"2504.18633","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SI-CLAD computes selective-inference p-values for DBSCAN-detected anomalies, controlling the false positive rate at a user-specified level while raising true detection rates.","lead":"This paper introduces SI-CLAD, a selective-inference method that produces valid p-values for anomalies flagged by the DBSCAN clustering algorithm, so the false positive rate stays below a chosen level. It is a step toward adding statistical guarantees to popular unsupervised anomaly detectors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 3's sign error makes every non-neighborhood constraint vacuous, so Z is a superset of the DBSCAN selection event and exact FPR control is not established.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing spot. Lemma 1 is standard SI conditioning, and Lemma 2 is a standard line parametrization; the entire novelty funnels through Lemma 3 and Algorithm 1. The sign error in Appendix A.3 is not cosmetic: with σ_ij=-1 the claimed inequality is identically true, so the over-conditioning region is not the set of z preserving the eps-neighborhoods. Because Algorithm 1 trusts Lemma 3 when appending [L_z,R_z] to Z, the final Z can include z values where DBSCAN does not yield O_obs. The p-value in Eq. (12) is then a truncated-normal tail probability over the wrong support, so uniform calibration under the null is unproven. This is an internal inconsistency, not a disagreement with consensus. The empirical FPR plots do not rescue the proof: simulations at n≤200 with 500 repetitions have limited resolution, and if the released code corrects the inequality, the paper text mis-describes the method actually evaluated. A corrected Lemma 3 would likely restore the structure of the argument, so conditional acceptance pending that correction remains the appropriate disposition; I therefore do not move the reader's verdict. The line-search exactness issue raised by the reader is real but secondary: even an exact line search over a wrongly computed Z_oc would inherit the error.","tokens_in":11708,"tokens_out":8706,"duration_ms":88889,"concrete_test":"Take n=2, a=(0,10), b=(1,0), eps=0.5, and consider pair (i,j)=(1,2). At z=10, point 2 is outside the eps-neighborhood of point 1, so σ_12=-1; Lemma 3 yields -|z-10|^2≤0.25, which holds for all z, whereas the true non-neighborhood condition is |z-10|^2>0.25. Brute-force enumerate z∈[-20,20] on a fine grid, computing N_eps(X_1(z)) exactly and comparing it with N_eps at z=10; the set of z where these sets agree is not all of [-20,20], while Lemma 3's interval is. This counterexample refutes Lemma 3. Then rerun the n=100, MinPts=5, eps=0.2 univariate FPR simulation (500 reps) using a corrected interval computation and compare the empirical FPR with the paper's reported SI-CLAD FPR at α=0.05.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Lemma 3 (Appendix A.3) is the step that turns the intractable DBSCAN selection event into the quadratic inequalities used in Algorithm 1. The paper defines σ_ij=+1 if j∈N_eps(X'_i) and σ_ij=-1 otherwise, then characterizes Z_oc(X') as ∩_{(i,j)} {z: σ_ij ||X_i(z)-X_j(z)||^2 ≤ eps^2}. For σ_ij=-1 this inequality becomes -||X_i(z)-X_j(z)||^2 ≤ eps^2, which is true for every z because the left side is nonpositive and the right side is positive; every 'must remain outside the eps-neighborhood' constraint drops out of the intersection, so Z_oc is a superset of the true over-conditioning region. The correct inequality in the same notation is σ_ij||X_i(z)-X_j(z)||^2 ≤ σ_ij eps^2, i.e., ||X_i(z)-X_j(z)||^2 ≥ eps^2 when σ_ij=-1. Algorithm 1 appends [L_z,R_z]=Z_oc(X(z)) to Z whenever the left endpoint z has O_X(z)=O_obs; if [L_z,R_z] is too wide, Z contains parameter values on which DBSCAN does not produce O_obs. Equation (12) then evaluates a truncated-normal tail probability over the wrong support. This is not automatically conservative: enlarging the truncation set changes the normalizing constant and can shift the p-value downward, inflating the FPR. Lemma 1's exact equality P(p_selective≤α)=α therefore does not follow from the written proof, and the claimed 'highest TPR among methods that control FPR' is not supported for the method as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes SI-CLAD, a selective-inference method that computes p-values for anomalies detected by DBSCAN. The data are modeled as Gaussian; for each detected anomaly j, the test statistic is the deviation from the mean of the remaining points, and the p-value is computed conditional on the DBSCAN output and on a nuisance component (Eqs. 3-8). The main technical step is Lemma 3, which claims that the set of z values along a one-dimensional parametrization preserving all eps-neighborhoods is described by quadratic inequalities, and Algorithm 1 uses that characterization in a line search to build the truncation region Z; the p-value is then a truncated-normal tail probability (Eq. 12). Lemma 1 asserts exact FPR control. Experiments on synthetic and real data report FPR control and high TPR, and code is available. Section 3.4 sketches a multi-dimensional extension.","tokens_in":12048,"tokens_out":8795,"duration_ms":86604,"significance":"If the derivation were sound, this would be a useful contribution to the growing selective-inference literature: it would be the first valid p-value method for DBSCAN-based anomaly detection, with an explicit algorithm and reproducible code, and the idea of over-conditioning on eps-neighborhoods to tame DBSCAN's discontinuous selection event is natural. The empirical FPR results are encouraging. However, the central theoretical claim currently rests on an incorrect characterization in Lemma 3 and on an unproven exactness claim for the line-search algorithm, so the paper cannot be accepted in its present form.","major_comments":[{"comment":"The inequality for non-neighbors is written incorrectly. With sigma_ij = -1, the condition sigma_ij ||X_i(z)-X_j(z)||^2 <= eps^2 becomes -||X_i(z)-X_j(z)||^2 <= eps^2, which holds for every z; all 'must remain outside' constraints are vacuous, so the computed Z_oc is a superset of the true over-conditioning region. The correct unified inequality in the same notation is sigma_ij ||X_i(z)-X_j(z)||^2 <= sigma_ij eps^2, i.e., ||X_i(z)-X_j(z)||^2 >= eps^2 when sigma_ij = -1. Because Algorithm 1 builds [L_z,R_z] from this flawed Z_oc, the truncation set used in Eq. (12) is not the actual selection event. Enlarging the truncation set changes the truncated-normal normalizing constant and is not automatically conservative; the p-value can move downward, so the equality P(p_selective <= alpha) = alpha in Lemma 1 is not established.","section":"Appendix A.3, Lemma 3"},{"comment":"The paper does not prove that the finite line search recovers the union in Eq. (14). The algorithm checks the DBSCAN output only at the left endpoint z, appends the whole interval [L_z,R_z], and then jumps to R_z + delta; intervals of the selection event shorter than delta, or lying between R_z and the next sampled point, are missed, and the choice of z_min and z_max is left unspecified. Since Eq. (12) requires the exact Z in Eq. (11), exact FPR control is not guaranteed for the implemented procedure. A proof of exact recovery, or a conservative modification with explicit handling of the grid and tolerance, is needed.","section":"Algorithm 1 / Eq. (14)"},{"comment":"The multi-dimensional extension is not derived. Gamma_j is a sum of absolute deviations; its representation as a linear form eta_j^T vec(X) fixes the sign vector s from the observed data, but the distributional statement in Eq. (16) conditions only on O_X and s_X. The induced truncation region for the signs is not characterized, and the claim that the techniques of Sections 3.2 and 3.3 apply 'straightforwardly' is unsupported. In particular, the unconditional distribution of Gamma_j under the null is a folded normal, not a normal, so the truncated-normal argument requires an explicit conditioning event and a proof.","section":"Section 3.4, Eqs. (15)-(16)"}],"minor_comments":[{"comment":"The statement that the Bonferroni adjustment factor 'scales exponentially with n, specifically reaching 2^n' is inaccurate; the standard correction for n hypotheses is alpha/n.","section":"Section 1, Related works"},{"comment":"There are several typos: 'parameterired' should be 'parameterized', 'modesl' should be 'models', and 'performace' should be 'performance'.","section":"Appendix A.3 and throughout"},{"comment":"Conditioning on the continuous variable Q_X = Q_obs and integrating over Q_obs with a density is informal; a measure-theoretic conditioning argument as in Lee et al. (2016) should be cited or sketched.","section":"Appendix A.1, proof of Lemma 1"},{"comment":"The notation Z is used both for the truncation region and for the random variable after Eq. (12); this is confusing and should be disambiguated.","section":"Section 3.2, Eq. (11)-(12)"},{"comment":"The FPR and TPR results are reported as point values without error bars or standard errors; given the paper's exactness claims, some uncertainty quantification would strengthen the empirical support.","section":"Section 4.1, Figures 3-6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a fixable but important technical error in the proof of Lemma 3, and the algorithm's exactness and multi-dimensional claims need substantial work. I would invite a revision rather than reject, because the core SI formulation and the empirical demonstration of FPR control are promising. The availability of code and a pip package is a further positive factor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is the first selective-inference method for DBSCAN-based anomaly detection, and the divide-and-conquer line search for tracing the truncation region is a genuine computational idea. The paper adapts Lee et al. (2016) and Duy–Takeuchi (2022) carefully to DBSCAN's selection event, and the experiments compare against sensible baselines. If the theory holds up after a fix, this fills a real gap: no one else has given valid p-values for density-based clustering outputs.\n\nThe stress-test concern is correct, and it's not a nitpick. In the proof of Lemma 3 (Appendix A.3), the constraint for non-neighbors is written as σ_ij ||Xi(z)−Xj(z)||^2 ≤ eps^2 with σ_ij = −1. That inequality is true for every z, so all the \"must stay outside the eps-neighborhood\" conditions drop out. The RHS should be σ_ij·eps^2, which for σ_ij = −1 reads ||·||^2 ≥ eps^2. As written, Z_oc is a superset of the true over-conditioning region. Then Algorithm 1 unions these intervals, so the final truncation set Z can contain z values on which DBSCAN does not actually produce O_obs. That means equation (12) evaluates the truncated-normal tail over the wrong support, and Lemma 1's exact equality does not follow. I would not lean on the vague \"it's conservative\" intuition—enlarging a truncation set is not automatically conservative for a two-sided p-value. So the paper's central theoretical claim, exact false-positive control, is currently unsupported.\n\nA second, smaller gap: Algorithm 1's finite line search is asserted to recover Z exactly, but there's no proof that the tolerance δ, the step rule, and the arbitrary z_min/z_max don't miss intervals. This is fixable but needs a formal statement. The multi-dimensional extension is a sketch, though that is minor and follows the same machinery.\n\nTo the authors' credit, the experiments show actual FPR control in the settings tested, and the code is available. That suggests the approach may work in practice, but 500 repetitions in a few settings is not a substitute for a correct proof.\n\nWho is this for? Anyone working on post-selection inference for unsupervised methods, and practitioners who want error control on DBSCAN output. It deserves a serious referee after the sign error is fixed and the line search is rigorously analyzed. I would send it to review, not desk reject—the kernel is sound and testable.\n\nMy recommendation: engage with it, but insist on a corrected Lemma 3 and an explicit theorem about what Algorithm 1 computes.","headline":"The idea—selective inference for DBSCAN-detected anomalies—is worth taking seriously, but the key lemma has a sign error that voids the stated FPR guarantee, so the paper needs a corrected proof before it can be trusted.","tokens_in":12610,"tokens_out":2069,"would_cite":false,"duration_ms":22679,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F03","62H30"],"pacs":[],"model":"deepseek-v4-flash","headline":"DBSCAN's anomaly flags can be tested with exact false-positive control, the paper proves.","keywords":["selective inference","anomaly detection","DBSCAN","false positive control","post-selection inference","clustering-based anomaly detection","truncated normal distribution","parametric programming"],"falsifier":"Run DBSCAN on a small null data set, enumerate by brute force every point along the conditioning line that yields the same anomaly set, and compare that set with the intervals returned by Algorithm 1; any discrepancy means the selective p-values are not the exact truncated-normal tail probabilities and the false positive rate at $\\alpha$ will not be exact. Equivalently, simulate many null data sets and check whether the empirical false positive rate of SI-CLAD exceeds $\\alpha$ by more than Monte Carlo error.","tokens_in":11465,"feed_emoji":"🎯","tokens_out":6615,"duration_ms":60344,"temperature":0.7,"pith_summary":"This paper asks whether anomalies flagged by a clustering-based detector can be tested statistically even though the same data were used to select them. It answers yes for DBSCAN: by conditioning on the exact event that DBSCAN returns the observed anomaly set, it constructs a selective p-value for each flagged point and proves that, under the null hypothesis, the p-value is uniform, so the probability of any false detection stays at the chosen significance level. The authors also give a line-search algorithm that traces the set of data sets yielding the same anomaly set, and they report experiments on synthetic, correlated, and real data where their method holds the false positive rate at the target level while keeping a higher true detection rate than existing valid alternatives.","feed_headline":"DBSCAN anomalies can now get exact p-values","feed_subtitle":"Conditioning on the detector's choice removes double-dipping bias and caps false alarms at any α.","key_machinery":"The load-bearing object is the truncation region $\\mathcal{Z}$, the set of scalar values $z$ along the line $X(z) = a + bz$ for which DBSCAN produces the same anomaly set as the observed data; the selective p-value is a tail probability of the truncated normal distribution of $|Z|$ over $\\mathcal{Z}$. Because $\\mathcal{Z}$ is hard to compute directly, the paper uses over-conditioning: it first conditions on each point's eps-neighborhood being unchanged, which Lemma 3 expresses as a system of quadratic inequalities in $z$. Then it stitches these regions together along the parametrized line with a divide-and-conquer line search, alternating DBSCAN runs and analytic interval updates to build $\\mathcal{Z}$ as a union of intervals. This machinery turns a discrete, combinatorial selection event into an interval computation, which is what makes an exact p-value tractable.","core_discovery":"The paper's central claim is that false detection control is attainable for DBSCAN-based anomaly detection. Given the observed data, SI-CLAD treats each flagged point as a test of whether its mean equals the mean of the non-flagged points. The test statistic is projected onto a scalar direction, and the p-value is computed from the truncated normal law of that statistic conditional on the event that DBSCAN returns the same anomaly set and on a nuisance component. Lemma 1 states that this selective p-value satisfies $P(p \\leq \\alpha) = \\alpha$ exactly under the null, not merely approximately, and the experiments show that the false positive rate is controlled across univariate, multidimensional, and correlated settings while the true positive rate is the highest among methods that are valid.","pith_inferences":["If the computed truncation region is indeed exact, the same line-search-plus-over-conditioning recipe should transfer to other density-based detectors such as OPTICS or DENCLUE whenever their selection events can be written as quadratic inequalities; the paper gestures at this extension but does not prove it.","The covariance is assumed known or estimated from independent data, so a natural stress test is to estimate it from the same data and measure how much the false positive rate inflates when the truncated normal law is misspecified.","For small sample sizes, comparing Algorithm 1's intervals against brute-force enumeration of DBSCAN-stable intervals would separate approximation error from statistical validity and could be used to choose the tolerance parameter.","Because the standard exponential multiple-testing correction becomes hopeless as the sample size grows, the practical value of conditioning the correction factor down to one grows precisely in the large-$n$ regime where naive tests fail."],"forward_implications":["An analyst using DBSCAN with pre-specified eps and MinPts can report a significance level for each flagged anomaly without correcting for the fact that the data determined the flags.","The same conditional test works in multiple dimensions and under correlated noise, so the guarantee is not limited to toy one-dimensional settings.","The over-conditioned version of the method is also valid, but the full line search recovers more power, so the gap between the two quantifies the price of over-conditioning.","Computational cost per p-value grows roughly linearly in sample size and dimension in the reported experiments, making the exact test feasible on moderate datasets.","For Euclidean-distance DBSCAN the method inherits the validity of the truncated-normal selective inference framework; for other distance functions the supplied characterization does not apply."],"supporting_citations":[{"why":"Supplies the selective inference template: conditioning on the selection event turns the p-value into a truncated-normal tail probability.","marker":"[Lee et al., 2016]"},{"why":"Defines DBSCAN, the clustering-based detector whose selection event SI-CLAD conditions on.","marker":"[Ester et al., 1996]"},{"why":"Provides the parametric programming and line-search technique used to trace the truncation region.","marker":"[Duy and Takeuchi, 2022]"},{"why":"Documents the double-dipping bias that motivates conditioning on the selection event.","marker":"[Kriegeskorte et al., 2009]"},{"why":"Gives the earlier valid-inference-after-outlier-removal setting that SI-CLAD extends to clustering-based anomaly detection.","marker":"[Chen and Bien, 2020]"}],"fun_headline_variants":["Exact p-values for DBSCAN anomalies","DBSCAN outlier tests now have exact p-values","Conditioning on DBSCAN selection gives exact p-values","Cluster anomaly detection gets exact false alarm control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire false-positive guarantee depends on the assumption that the algorithm computes exactly the set of data sets on which DBSCAN would flag the same anomalies, and the paper assumes both the quadratic characterization and the finite line search recover that set rather than a superset or approximation of it.","fun_headline_variants_meta":{"raw":{"variants":["Exact p-values for DBSCAN anomalies","DBSCAN outlier tests now have exact p-values","Conditioning on DBSCAN selection gives exact p-values","Cluster anomaly detection gets exact false alarm control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000278,"raw_usage":{"total_tokens":1614,"prompt_tokens":866,"completion_tokens":748,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":687}},"tokens_in":482,"tokens_out":748,"duration_ms":7662,"temperature":1.0,"reasoning_tokens":687,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:13:53.410095+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DBSCAN on a small null data set, enumerate by brute force every point along the conditioning line that yields the same anomaly set, and compare that set with the intervals returned by Algorithm 1; any discrepancy means the selective p-values are not the exact truncated-normal tail probabilities and the false positive rate at $\\alpha$ will not be exact. Equivalently, simulate many null data sets and check whether the empirical false positive rate of SI-CLAD exceeds $\\alpha$ by more than Monte Carlo error.","supporting_citations":[],"review_version":1}