{"id":"9535792f-13bf-4ab4-9878-459196015e67","arxiv_id":"2506.14108","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"β-integrated local depth averages local depth over all locality levels, and its partitioned matrix representation gives interpretable local centrality scores that improve depth-based classification and outlier detection.","lead":"The paper introduces β-integrated local depth, a depth score that averages a point's local centrality over many neighborhood sizes, smoothing the erratic behavior of choosing a single locality level. It also defines a partitioned matrix showing how each point contributes to another's local depth, and applies both to classification and outlier detection.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Property 2's continuity proof uses unverified strict-monotonicity/convexity conditions and a false Hausdorff-convergence implication; the inherited-continuity claim is not established.","rationale":"I read the paper as making two intertwined claims: (i) β-ILD is a smoothed local depth that inherits useful theoretical properties, including continuity in x, and (ii) it yields practical gains in classification and outlier detection. The theoretical inheritance claim is load-bearing because the applications motivate the definition and because continuity in x is repeatedly used to justify the measure as a stable local centrality index. The reader's weakest-assumption analysis identified a genuine gap in Property 2: the proof invokes Dyckerhoff [38] under conditions that are stronger than quasi-concavity and are not verified for the family of symmetrized distributions Q_{x_n}. In addition, the proof's transition from Hausdorff convergence of central regions to pointwise convergence of indicator functions is not valid as stated; Hausdorff convergence only gives closeness, not eventual membership, so the DCT argument requires a separate boundary-measure argument or a different mode of set convergence. This is a proof gap rather than a demonstrated counterexample, so the appropriate disposition is the reader's conditional acceptance: the property may be true, but it is not established by the submitted argument. I do not see a reason to move the verdict, because the gap is localized and potentially patchable, and the empirical concerns, while real, are secondary to the theoretical load-bearing point. I therefore agree with the reader's weakest-assumption identification and leave the verdict unchanged.","tokens_in":19315,"tokens_out":15890,"duration_ms":181007,"concrete_test":"Independently verify the implication used in the Appendix proof of Property 2: does Hausdorff convergence of central regions R^β_{x_n} → R^β_x imply weak convergence of the conditional distributions P^β_{x_n} → P^β_x? Concretely, test the proof's intermediate claim that every z ∈ R^β_x is eventually in R^β_{x_n} using A = [0,1] and A_n = [1/n, 1+1/n]; since this claim is false even for convex sets, determine whether a.e. convergence of indicators can be recovered from convexity plus μ(∂R^β_x) = 0. If not, the DCT step in Property 2 fails; if yes, state explicitly which additional boundary or strict-monotonicity assumption is needed and patch the proof.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing gap is in Property 2 (continuity of LDβ and ILD in x). The proof's Step 2 asserts that, by Dyckerhoff [38], strict monotonicity plus convexity of central regions implies Hausdorff convergence R^β_{x_n} → R^β_x. But the stated hypotheses of Property 2 are only quasi-concavity plus Assumption 1: quasi-concavity gives convex upper level sets but not strict monotonicity, and no verification is given that D(·|Q_{x_n}) satisfies the additional Dyckerhoff conditions. The proof then uses a stronger false implication: Hausdorff convergence does not imply that every z ∈ R^β_x is eventually in R^β_{x_n} (e.g., intervals [1/n, 1+1/n] converge to [0,1] in Hausdorff distance, yet 0 is in no approximating interval). The subsequent DCT step requires a.e. convergence of indicators 1_{R^β_{x_n}} → 1_{R^β_x}, which needs control of the boundary measure μ(∂R^β_x) = 0 and is not established. Since continuity of ILD in x is one of the advertised inherited properties, and Property 4 also relies on continuity of LDβ in β, the central theoretical claim is conditional on an unproven argument. The empirical 'significant improvements' claim would also need error bars and non-oracle parameter selection, but the proof gap is the more fundamental issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes β-integrated local depth (β-ILD), defined as ILD(x|P_X,W)=∫_0^1 LD(β|x,P_X)dW(β), where LD is the Paindaveine–Van Bever β-local depth. The authors argue that integrating over the locality parameter smooths the unstable sample β-LD function (Proposition 2.1), and that the resulting functional inherits consistency, T-invariance, continuity in x, vanishing at infinity, and an extreme-locality centrality property. They introduce a paired/partitioned matrix PILD whose row sums equal β-ILD, and they report simulation studies for classification and outlier detection that are claimed to show significant improvements over global and local depth methods.","tokens_in":19663,"tokens_out":8170,"duration_ms":93139,"significance":"If the theoretical claims are established, β-ILD would be a useful tool for removing the sensitivity of local depth to the locality parameter, and the PILD matrix provides a genuinely interpretable local-centrality summary. The paper has clear strengths: Definition 2.4 is simple and natural; Proposition 2.1 gives a clean, parameter-free smoothing bound; the strategy of transferring consistency and invariance properties from β-LD to β-ILD by dominated convergence is sound in outline; and the simulation setups are described in enough detail to be reproducible. However, the continuity-in-x proof rests on an unverified and partly false central-region convergence argument, and Property 4 has a normalization problem in its mean-value step. These gaps affect advertised inherited properties, so the theoretical core of the paper is not yet established in its current form.","major_comments":[{"comment":"The proof of continuity in x is not valid under the stated assumptions. Step 2 invokes Dyckerhoff [38] to assert Hausdorff convergence of the central regions R^β_{x_n} to R^β_x, but that result requires strict monotonicity and convexity of central regions, whereas Property 2 only assumes quasi-concavity plus Assumption 1. Quasi-concavity gives convex superlevel sets, not strict monotonicity, and no verification is supplied that the symmetrized mixture distributions satisfy the extra Dyckerhoff conditions. More seriously, the proof then uses the implication that Hausdorff convergence R^β_{x_n} → R^β_x entails eventual inclusion of every z∈R^β_x in R^β_{x_n}. This implication is false: the intervals [1/n,1+1/n] converge to [0,1] in Hausdorff distance, yet 0 belongs to none of the approximating intervals. Consequently, the claimed pointwise convergence of indicators 1_{R^β_{x_n}} → 1_{R^β_x} is not justified, and the subsequent dominated-convergence step also requires a boundary condition μ(∂R^β_x)=0 that is not established. Since continuity of ILD in x is one of the advertised inherited properties, this is a load-bearing gap; it can likely be repaired by adding strict-monotonicity/convexity and boundary-measure-zero assumptions, but as written Property 2 is not proven.","section":"§3, Property 2 (Appendix proof)"},{"comment":"The proof of the extreme-locality centrality property appears to misuse the weighted mean value theorem. In the statement, W is fixed as a probability measure on (0,1], and ILD_B(x|P)=∫_0^B LD(β|x,P)dW(β). The proof then applies the mean value theorem as if ∫_0^B w(β)dβ=1 for every B, but for a fixed W this integral tends to 0 as B→0. The sentence \"where W(0,B]=1\" appears to redefine W for each B; if ILD_B is instead defined with W conditionally normalized to (0,B], this is a different object from the original integrated depth and the notation should say so explicitly. Without such normalization, the displayed mean-value identity is incorrect, and the conclusion lim_{B→0} ILD_B(x|P)=lim_{β→0} LD_β(x|P) does not follow from the argument given.","section":"§3, Property 4"},{"comment":"The empirical claims of \"significant improvements\" are not supported by uncertainty quantification or by the stated parameter-selection protocol. In Table 1, the text states that PILD and B-PILD are \"significantly better\" in some setups and that certain methods \"consistently outperform\" others, but no standard errors, confidence intervals, or significance tests are reported across the 100 simulations. In Table 2, the authors say \"we select the optimal parameters that maximize precision\" for all methods; since this selection is performed on the same data used for evaluation, the reported precision values are oracle-tuned and cannot be compared directly as estimates of unsupervised out-of-sample performance. The abstract's claim of significant improvements should either be accompanied by proper error bars and non-oracle tuning, or be qualified accordingly.","section":"§4.1–4.2, Tables 1–2"}],"minor_comments":[{"comment":"The β-neighborhood is defined as \"the set of ⌈nβ⌉ points with highest depth values\", but ties in depth values are not addressed; a tie-breaking rule or a deterministic convention is needed for the sample β-LD and PILD to be well defined.","section":"Definition 2.3"},{"comment":"Definition 2.4 allows an arbitrary probability measure W, but the sample formula in Eq. (7) and the theoretical results in Section 3 assume an absolutely continuous W with density w. Please state explicitly where absolute continuity is required and what happens for discrete W.","section":"Eq. (7) and Section 3"},{"comment":"Several entries in the WDBC row are run together as \"0.60.70.2\" and similar strings, making the table unreadable; the entries need proper spacing or column separation.","section":"Table 2"},{"comment":"There are several typographical errors, including \"multivatiate\" in the introduction and \"varing\" in the conclusion; a careful proofread is needed.","section":"Throughout"},{"comment":"The phrase \"the probability measure W remains unchanged under any transformation of P\" is imprecise; since W is a measure on (0,1], it does not depend on P, so the intended condition should be stated more clearly, for example that W is the same measure in both ILD(T(x)|P_{T(X)}) and ILD(x|P_X).","section":"Property 3"}],"recommendation":"major_revision","confidential_remarks":"The core idea is sound and the smoothing bound in Proposition 2.1 is a genuine contribution, but the continuity proof in Property 2 needs a substantial repair, and Property 4 has a normalization issue. I would be willing to see a revision that either proves these properties under clearly stated additional assumptions or removes them from the list of claimed inherited properties. The empirical sections also need standard errors and a non-oracle parameter-selection protocol before the paper can claim significant improvements."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The β-ILD definition is a genuine generalization of Paindaveine and Van Bever's local depth: integrating over the locality parameter β is different from integrating over directions as in Fernandez-Piana and Svarc, and it's a natural way to avoid choosing β. The PILD matrix is new and gives an interpretable decomposition of local depth; the row-sum identity is nice. Proposition 2.1's smoothing bound is clean and the proof is correct. Consistency, T-invariance, and vanishing at infinity mostly follow by DCT or direct arguments. The paper does a good job of making the intuition visual, including the Shiny app.\n\nThe problems are concentrated in Property 2 and the empirical claims. Property 2's proof assumes, via Dyckerhoff, that central regions converge in Hausdorff distance under strict monotonicity and convexity, but the stated hypothesis is only quasi-concavity. No verification is given that the depth functions satisfying Assumption 1 meet the stronger Dyckerhoff conditions. Worse, the proof uses Hausdorff convergence to assert that every point in the limit region is eventually in the approximating regions; that implication is false (e.g., [1/n, 1+1/n] → [0,1] in Hausdorff). The DCT step needs boundary measure zero, which is not established. So continuity of ILD in x is not proven as stated. That's a real gap in one of the advertised properties.\n\nOn the applications: the classification boxplots show the expected patterns, but 'significant improvements' is not supported by error bars or tests. In outlier detection, parameters are selected by maximizing precision on the same data (oracle), so the numbers are optimistic. The empirical section is suggestive, not demonstrative.\n\nWho's this for? Anyone working on depth-based methods and local centrality. The definition and matrix are worth knowing about, and the proof gap is a good exercise for a referee. I'd send it to a serious referee rather than desk reject; the problems are fixable and the core idea is sound. My recommendation: engage with it, but require the authors to fix Property 2 and temper the empirical language.","headline":"A genuinely new local depth construction and a useful PILD matrix, but the continuity proof has a gap and the empirical claims run ahead of the evidence.","tokens_in":20156,"tokens_out":2919,"would_cite":true,"duration_ms":31239,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","62G20","62H30"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes β-integrated local depth, a weighted integral of β-local depth that smooths fluctuations, keeps global-depth properties, and via its PILD decomposition improves depth classifiers and outlier detection.","keywords":["β-integrated local depth","local depth","data depth","partitioned local depth","PILD matrix","classification","outlier detection","depth-based classifier"],"falsifier":"Construct an absolutely continuous distribution and a quasi-concave depth satisfying Assumption 1 for which the central regions $R_\\beta^x$ do not converge in Hausdorff distance as $x_n\\to x$ (e.g., a depth with non-strictly-monotone or non-convex level sets), and check whether the corresponding β-LD is continuous in x; a discontinuity there would invalidate the proof of Property 2 and the paper's continuity claim for ILD.","tokens_in":19135,"feed_emoji":"📊","tokens_out":13902,"duration_ms":116168,"temperature":0.7,"pith_summary":"This paper proposes β-integrated local depth (β-ILD), the weighted integral of β-local depth with respect to the locality parameter β, and claims that this smoothing makes local depth less sensitive to the choice of β while preserving the structural guarantees of global depth: consistency, T-invariance, continuity in x, and vanishing at infinity. A discrete partition of the integral yields the PILD matrix, whose entries record how much each point contributes to the local depth of every other point and whose row sums recover β-ILD; column sums are then an interpretable local-centrality score. The paper argues, and demonstrates on simulations and benchmark datasets, that classifiers and outlier detectors built on β-ILD and PILD outperform their unintegrated depth-based counterparts, especially for non-convex or multimodal distributions. A curious reader should care because this offers a principled way to remove (or reduce) the tuning-parameter burden of local depth and to turn depth values into pairwise influence scores usable by other algorithms.","feed_headline":"Integrating over beta stabilizes local depth and improves classifiers","feed_subtitle":"A weighted average over all locality levels yields a stable centrality score and a point-to-point contribution matrix.","key_machinery":"The load-bearing object is the integrated local depth $ILD(x|P_X,W)=\\int_0^1 LD(\\beta|x,P_X)\\,dW(\\beta)$ (Definition 2.4), with $W$ a probability measure on $(0,1]$, plus its discrete partition $PILD(z|x,X)=\\sum_{i=1}^b \\mathbf{1}_{\\{z\\in N_{\\beta_i}^x\\}} \\frac{LD(\\beta_i|x,X)}{\\lceil n\\beta_i\\rceil} \\int_{\\beta_{i-1}}^{\\beta_i} w(\\beta)\\,d\\beta$ (Eq. 11). The integration smooths the local-depth curve; the bound of Proposition 2.1 (max adjacent fluctuation of ILD ≤ $\\Delta/2$, where $\\Delta$ is the max jump of LD) is the quantitative guarantee that this smoothing works. The partitioning then converts the scalar depth into a pairwise contribution matrix whose row sums equal ILD, so that column sums—how much a point is leaned on by others, weighted by their local depth—serve as the paper's new interpretable centrality measure.","core_discovery":"The central claim is that the map $\\beta\\mapsto LD(\\beta|x,P_X)$, which can jump substantially between neighbouring $\\beta$ levels, can be profitably replaced by $ILD(x|P_X,W)=\\int_0^1 LD(\\beta|x,P_X)\\,dW(\\beta)$. Proposition 2.1 shows that the adjacent fluctuation of this integrated version is at most half the maximum jump of the unintegrated local depth, so integration genuinely stabilizes the measure. The paper then shows that when β-ILD is computed in discrete steps and each point's share of the neighborhood is tracked, the resulting PILD matrix decomposes each point's local depth into pairwise contributions, with row sums exactly equal to ILD and column sums acting as a local-centrality score. On the theoretical side, β-ILD inherits consistency, T-invariance, continuity in x, centrality in the extreme-locality limit, and vanishing at infinity from the underlying depth; on the applied side, the paper exhibits classifiers and outlier scores based on these objects that improve over standard depth-based algorithms across several setups.","pith_inferences":["The same integration-over-tuning-parameter recipe could stabilize other local depth families, such as kernelized spatial depth over bandwidth σ or lens depth over radius, with an analogous bound whenever the family is bounded and stepwise continuous in its parameter.","The PILD matrix is effectively a directed, weighted graph on the sample with row sums equal to a global centrality score; normalizing and symmetrizing it (as done for LOF) suggests it could seed spectral clustering, community detection, or low-dimensional embedding, none of which the paper explores.","A distinction the paper does not draw: a point's column sum in PILD need not align with its own ILD, so a point with low ILD can still be a strong contributor to others' local depth; testing whether such 'influential non-central' points behave differently in clustering or outlier detection would be a natural follow-up.","If the Hausdorff-convergence premise fails for some depth, the continuity claim would need repair rather than the whole construction, since the integral in β may still smooth out isolated discontinuities in x—an extension worth checking."],"forward_implications":["With a uniform weight over $(0,1]$, the parameter-free Full-ILD classifier outperforms the max-depth classifier in most of the paper's classification setups, and the cross-validated B-ILD and B-PILD variants achieve the best average ranks among all tested classifiers.","Column sums of the PILD matrix give an outlier score that beats global depth, β-LD, and β-ILD on the benchmark datasets; the PILD-based similarity matrix also improves LOF on WPBC, WDBC, and SpamBase.","The PILD matrix provides a point-to-point decomposition of local depth where each row sums to β-ILD, making it a depth-based counterpart of the PaLD matrix used in local community depth.","β-ILD retains T-invariance, consistency, and vanishing at infinity automatically when the underlying depth does, so the integration does not sacrifice the structural guarantees that make depth functions attractive.","Proposition 2.1's half-jump bound means the choice of locality level matters less for β-ILD than for β-LD, which is the theoretical basis for the method's improved stability."],"supporting_citations":[{"why":"Supplies the β-local depth definition, its consistency theorem, and the extreme-locality centrality result that β-ILD directly extends.","marker":"[1]"},{"why":"Gives the axiomatic depth properties (P1–P4) used as Assumption 1 and the Type A/B/D classification invoked in Proposition 3.1.","marker":"[5]"},{"why":"Provides the central-region convergence in Hausdorff distance used in the proof of continuity of β-LD (and hence β-ILD) in x.","marker":"[38]"},{"why":"Introduces the PaLD/LCD framework whose partitioned-decomposition idea is adapted to build the PILD matrix.","marker":"[13]"},{"why":"Supplies depth-based classifiers and the D-knn baseline, and the half-moon classification setup reused as Setup 2.","marker":"[20]"},{"why":"Defines the unsupervised outlier-detection evaluation, benchmark datasets, and precision metric used in Section 4.2.","marker":"[29]"},{"why":"Formalizes the max-depth classifier that is the principal baseline for the classification experiments.","marker":"[22]"},{"why":"Introduces the max-depth classification principle that the proposed max-ILD and PILD classifiers generalize.","marker":"[27]"}],"fun_headline_variants":["Integrating local depth over beta stabilizes and improves classifiers","New β-integrated local depth yields stable centrality and better classifiers","Partitioned depth matrix improves outlier detection and classification","Local depth integration yields stable scores and contribution matrix","Taming local depth jumps with integration sharpens outlier scores"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The continuity of β-ILD in x is proved under the extra, unverified assumption that central regions of the symmetrized distributions converge in the Hausdorff sense as x varies, which needs strict monotonicity and convexity of those regions; if that convergence fails for some depth satisfying the stated assumptions, the continuity claim is not established.","fun_headline_variants_meta":{"raw":{"variants":["Integrating local depth over beta stabilizes and improves classifiers","New β-integrated local depth yields stable centrality and better classifiers","Partitioned depth matrix improves outlier detection and classification","Local depth integration yields stable scores and contribution matrix","Taming local depth jumps with integration sharpens outlier scores"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000892,"raw_usage":{"total_tokens":3810,"prompt_tokens":873,"completion_tokens":2937,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":2858}},"tokens_in":489,"tokens_out":2937,"duration_ms":20109,"temperature":1.0,"reasoning_tokens":2858,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:18:40.663815+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct an absolutely continuous distribution and a quasi-concave depth satisfying Assumption 1 for which the central regions $R_\\beta^x$ do not converge in Hausdorff distance as $x_n\\to x$ (e.g., a depth with non-strictly-monotone or non-convex level sets), and check whether the corresponding β-LD is continuous in x; a discontinuity there would invalidate the proof of Property 2 and the paper's continuity claim for ILD.","supporting_citations":[{"cited_title":"Paindaveine, G","cited_arxiv_id":null,"evidence_quote":"Supplies the β-local depth definition, its consistency theorem, and the extreme-locality centrality result that β-ILD directly extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the axiomatic depth properties (P1–P4) used as Assumption 1 and the Type A/B/D classification invoked in Proposition 3.1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the PaLD/LCD framework whose partitioned-decomposition idea is adapted to build the PILD matrix."},{"cited_title":"Paindaveine, G","cited_arxiv_id":null,"evidence_quote":"Supplies depth-based classifiers and the D-knn baseline, and the half-moon classification setup reused as Setup 2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the unsupervised outlier-detection evaluation, benchmark datasets, and precision metric used in Section 4.2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Formalizes the max-depth classifier that is the principal baseline for the classification experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the max-depth classification principle that the proposed max-ILD and PILD classifiers generalize."}],"review_version":1}