Pith. sign in

REVIEW 2 major objections 3 minor 13 references

The leave-one-out density influence score for outlier detection has an exact closed-form update and provably different asymptotic rates for inliers versus outliers.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A closed-form leave-one-out density influence score for the LBFP estimator separates inliers from outliers asymptotically and runs in time comparable to kNN or iForest.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Useful closed-form leave-one-out score for LBFP, but the claimed asymptotic separation is not actually proven: Proposition 2 needs an extra condition like ε_n g_n = o_p(1) to give distinct rates. the 2 major comments →

arxiv 2607.14335 v1 pith:D5WTNB7R submitted 2026-07-15 stat.ME

A Leave-One-Out Influence Statistic for Density-Based Outlier Detection

Aur\'elien Nicosia , Thierry Duchesne , Michel Carbon This is my paper
classification stat.ME MSC 62G0762G2062G35
keywords outlier detectionleave-one-outdensity-based scoringlinear blend frequency polygoncontamination modelasymptotic separationclosed-form updaterobustness
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes a score for detecting outliers by measuring how much the fitted density at each observation drops when that observation is removed. Using the Linear-Blend Frequency Polygon (LBFP) estimator, the authors show this leave-one-out score has an exact closed-form update, so computing it for all n points costs essentially one density fit rather than n refits. Under a Huber-style contamination model, they prove that the score for ordinary inliers decays as O_p((nV_b)^-1), while for contaminated observations the score times nV_b epsilon_n g_n(X_i) stays bounded, giving a formal asymptotic separation between the two classes. The practical claim is that this yields a density-based outlier score that is fast, has a direct probabilistic reading, and performs competitively with distance- and isolation-based benchmarks on simulations and on a 284,807-row, 29-variable credit-card fraud dataset.

Core claim

The central discovery is that the leave-one-out density influence score D_i = 1 - fhat_b,(-i)(X_i)/fhat_b(X_i) admits an exact algebraic update for LBFP estimators. Because the LBFP is a normalized sum of per-observation contributions on a fixed grid with fixed bandwidth, deleting X_i changes the fitted density at X_i only through the observation's own interpolation weight W_i, which lies in [2^-d, 1]. The resulting identity D_i = 1/(n-1)( W_i/(V_b fhat_b(X_i)) - 1 ) turns an n-fold refitting problem into a single full-sample fit. Proposition 2 then shows under a Huber contamination model that inliers have D_i = O_p((nV_b)^-1), while for outliers D_i n V_b epsilon_n g_n(X_i) = O_p(1); hence

What carries the argument

The Linear-Blend Frequency Polygon (LBFP) estimator, a binned density estimator that linearly interpolates multivariate histogram counts on a fixed rectangular grid of cells with volume V_b. Its contribution to the paper's argument is that its fitted value at any point is a normalized sum of per-observation indicator contributions with fixed interpolation weights, so the leave-one-out update is a simple correction: subtracting the observation's own weight W_i and renormalizing by n-1. The weight W_i, a product of per-coordinate max{u, 1-u} terms in [2^-d, 1], measures where the point sits inside its interpolation cell; it carries the geometric part of the score, while V_b fhat_b(X_i) carries

Load-bearing premise

The result rests on Assumption 2: at every contaminating observation, the fitted density is positive with probability tending to one and is not asymptotically smaller than the contamination contribution epsilon_n g_n(X_i); the paper asserts this as a sufficient condition rather than proving that the LBFP estimator always satisfies it, and Section 5 concedes sensitivity to bandwidth and dimension.

What would settle it

Construct a contamination model in which an outlier lies in a region where the fitted density is much larger than the contamination contribution, e.g., many inliers nearby and lambda_n(x) large; then check whether the outlier's D_i decays at the inlier rate O_p((nV_b)^-1) rather than satisfying D_i nV_b epsilon_n g_n(X_i)=O_p(1). Alternatively, test Assumption 2 directly by checking whether epsilon_n g_n(X_i)/fhat_b(X_i) = O_p(1) holds at outliers for the LBFP under the mixture model; a violation would destroy the claimed separation.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • All n leave-one-out scores are obtained from one LBFP fit plus an O(1) correction per point, removing the computational bottleneck that has limited leave-one-out density methods to small samples.
  • In the sparse-local-support regime (expected contaminating neighbors tending to zero), outlier scores are near one while inlier scores decay to zero, so a fixed threshold separates the classes with probability tending to one.
  • In the finite-local-support regime, outlier scores remain non-degenerate as n grows while inlier scores vanish, so ranking-based outlier identification (e.g., ROC-AUC) should improve with sample size.
  • The score inherits the LBFP's convergence theory for weakly dependent data, so the same closed-form update applies to spatial or temporal outlier detection under alpha-mixing dependence.
  • On the credit-card fraud data, the method reported an ROC-AUC of 0.901, a lift of 21.18 at the fraud rate, and a runtime of about 91 seconds on 284,807 transactions in 29 variables, suggesting the score is usable at scale.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The algebraic identity at the heart of Proposition 1 (any estimator that is a normalized sum of fixed per-observation contributions) suggests the same one-pass leave-one-out update could be derived for other linear smoothers, such as kernel density estimates with precomputed weights, not just the LBFP.
  • The score is essentially a finite-sample self-influence or leverage measure: it asks how much an observation props up its own fitted density. This connects outlier detection to classical influence diagnostics and could make the method's thresholds interpretable in terms of local leverage rather than purely geometric isolation.
  • The separation between classes hinges on the expected local contamination count lambda_n(x) = n epsilon_n P(Z in A_b(x)); an immediate practical extension is to choose bin widths to keep lambda small in candidate outlier regions, turning the asymptotic regimes into a bandwidth-selection rule.
  • The paper's own conclusion notes the method is sensitive to bandwidth choice and breaks down in dense high-dimensional data; an adaptive, locally varying bandwidth is the natural next step, and it is testable whether Proposition 2's separation survives such a modification.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 3 minor

Summary. The paper proposes a leave-one-out density influence score (LOODIS), D_i = 1 - fhat_{(-i)}(X_i)/fhat(X_i), and specializes it to the Linear-Blend Frequency Polygon (LBFP) estimator. Proposition 1 gives an exact closed-form expression for D_i, avoiding n leave-one-out refits. Proposition 2 states asymptotic rates under a Huber-type contamination model: inliers have D_i = O_p((nV_b)^{-1}) and outliers satisfy D_i n V_b ε_n g_n(X_i) = O_p(1). These rates are interpreted as implying distinct asymptotic orders and well-separated score distributions. Simulations over four contamination scenarios and a credit-card fraud application compare LOO-LBFP with LOO-KDE, LOF, KDEOS, kNN, and iForest, emphasizing computational speed.

Significance. The exact leave-one-out update is a genuine and useful result; Proposition 1 is cleanly proved and correctly identifies the LBFP interpolation weight. The computational advantage is convincingly documented (Tables 3 and 4, Figure 3). The simulation suite is broad, and the credit-card benchmark is reported without overclaiming superiority. The main unresolved issue is that the asymptotic separation advertised in the abstract and Section 5 is not a consequence of Proposition 2 as stated: the theorem gives rates but no lower bounds, and the condition under which the two rates differ is not part of the assumptions. This is a load-bearing theoretical claim, so the paper currently overstates its formal support. If the authors add the missing condition and either prove a lower-bound separation result or qualify the conclusion, the contribution would be worthwhile.

major comments (2)
  1. [§2.2, Proposition 2] The abstract and Section 5 claim that inliers and outliers have 'distinct asymptotic orders.' Proposition 2 as stated does not imply this. From the outlier bound D_i = O_p(1/(nV_b ε_n g_n(X_i))) and the inlier bound D_i = O_p((nV_b)^{-1}), the two orders differ only when ε_n g_n(X_i) = o_p(1). Assumption 1 only requires V_b ε_n g_n(X_i) = O_p(1); since V_b → 0, this permits ε_n g_n(X_i) to be O_p(1) or to diverge, in which case the outlier bound is no larger than the inlier bound. The informal discussion after Proposition 2 explicitly conditions on ε_n g_n(X_i) = o(1) in the λ→∞ case, but this condition never appears in Assumptions 1–2 or in the proposition. Please add the missing condition to Proposition 2, or restrict the claimed separation to the regimes in which it is actually proved.
  2. [§2.2 and §5] Proposition 2 gives only an upper bound for the outlier score; it does not provide a lower bound or a non-degeneracy statement. Thus even when ε_n g_n(X_i) = o_p(1), the theorem does not prove that outlier scores dominate inlier scores in the ranking sense required for detection. The λ-regime discussion in §2.2 says that the outlier distribution 'should still be non-degenerate' (λ finite) and 'can be large' (λ→0), which is heuristic. Section 5 states that the paper 'formally proved that the distribution of the LOO-LBFP score was well separated,' which is stronger than what the proof establishes. To make the central claim rigorous, the authors need either a lower bound on D_i for outliers that exceeds the inlier rate with probability tending to one (for the relevant λ regimes), or they should state the separation claim as an empirical observation rather than a formal consequence of Propos
minor comments (3)
  1. [§2.3] The bin-width rule is described as 'proportional to σ̂_rob n^{-1/(d+4)}', but the proportionality constant is not given in the main text. Please state the constant or refer explicitly to the supplementary calibration (S1).
  2. [Table 1] In the Ring scenario, the within-class means are 0.0057 (inliers) and 0.0076 (outliers) with standard deviations 0.0001 and 0.0007; this is not visually 'well separated.' Reporting an effect-size or overlap measure would help readers calibrate the claim.
  3. [§2.1] There is a typo at the start of §2.1: 'LetX_i' should read 'Let X_i'. Also, the proof of Proposition 1 could state explicitly that the grid and bin width are fixed by the definition of the score, not merely in the proof.

Circularity Check

0 steps flagged

No significant circularity: central LOO-LBFP identity is a direct algebraic consequence of the estimator definition; self-citations are present but not load-bearing.

full rationale

The central closed-form update (Proposition 1) is derived in Appendix A from the definition of the LBFP as a normalized sum of individual contributions, so D_i = 1/(n-1)(W_i/(V_b fhat_b(X_i))-1) is an identity, not a fitted prediction. Proposition 2 is a conditional statement under the explicitly stated sufficient Assumptions 1 and 2; the outlier conclusion D_i nV_b eps_n g_n(X_i)=O_p(1) follows by multiplying the Proposition 1 identity by nV_b eps_n g_n(X_i) and applying the assumed bounds. No parameter is fitted to a subset and then renamed a prediction, and no conclusion is assumed in the definition of the score. The paper cites the authors' prior work (Carbon & Duchesne 2024, 2025) for LBFP convergence properties and for the n^{-1/(d+4)} bin-width motivation, but those citations support the choice of estimator and practical bin-width rule rather than the derivation of Proposition 1 or the conditional rates in Proposition 2; they are not used as a uniqueness theorem or to forbid alternatives. The heuristic lambda-regime discussion in Section 2.2 and the 'well separated' wording in Section 5 may overstate what Proposition 2 alone guarantees (a point relevant to correctness), but this is a logical gap about whether eps_n g_n(X_i)=o_p(1), not circularity. Overall, the claimed derivation chain is self-contained given its assumptions; the only mild concern is non-load-bearing self-citation, hence score 2.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The method introduces no new physical or probabilistic entities. The free parameter is the implicit bin-width proportionality constant. The main burden is Assumption 2, which is a stated but unproved sufficient condition for the central separation result. The paper also relies on prior self-cited LBFP consistency results for the clean-data setting.

free parameters (1)
  • LBFP bin-width proportionality constant = unspecified (b_s is taken proportional to sigma_s^rob n^{-1/(d+4)})
    The bin-width rule in Section 2.3 has an implicit proportionality constant that is not given in the main text; it controls the effective smoothing and therefore which observations receive high scores.
axioms (5)
  • domain assumption LBFP pointwise consistency for the contaminated mixture density f_n with fixed grid and bandwidth (Assumption 2, inlier part)
    The paper imports consistency from the authors' earlier clean-data LBFP theory (Carbon & Duchesne 2024, 2025); the extension to the contaminated mixture f_n is assumed rather than proved.
  • ad hoc to paper For contaminating observations, epsilon_n g_n(X_i)/f_b(X_i) = O_p(1) and f_b(X_i) > 0 with probability tending to one (Assumption 2, outlier part)
    This sufficient condition keeps outlier scores from vanishing at the inlier rate; no theorem in the paper verifies it for the LBFP estimator under the contamination model.
  • domain assumption Inlier density bounded away from zero and contamination scale bounded: V_b epsilon_n g_n(X_i) = O_p(1) (Assumption 1)
    Prevents the denominator in the closed-form D_i expression from vanishing and bounds the local contamination contribution at outlier locations.
  • standard math Standard bin-width conditions V_b -> 0 and n V_b -> infinity with d fixed
    Required for the bin-count asymptotics and the rates in Proposition 2.
  • standard math Each observation contributes a positive amount to its own LBFP fitted value, with weight at least 2^{-d}/(n V_b)
    Follows from the definition of the LBFP interpolation weights; ensures the full-sample fitted density at X_i is never zero, so D_i is well defined.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of A Leave-One-Out Influence Statistic for Density-Based Outlier Detection." pith.science (2026). https://pith.science/paper/D5WTNB7R

@misc{pith2026260714335,
  author       = {Pith},
  title        = {Pith review of: A Leave-One-Out Influence Statistic for Density-Based Outlier Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D5WTNB7R}},
  note         = {Machine review of arXiv:2607.14335}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We propose a density-based leave-one-out influence score for unsupervised outlier detection. The motivation is that outliers are naturally associated with regions of very small probability density, but direct leave-one-out density refitting can be computationally prohibitive. We use the Linear-Blend Frequency Polygon (LBFP) estimator and define a score that compares the full-sample fitted density at an observation with the fitted density obtained after removing that observation, while keeping the grid and bandwidth fixed. The resulting statistic measures a relative density perturbation at the observation's own location. For the LBFP estimator, this score has an exact closed-form update, so the density estimator does not need to be refitted for each observation. This preserves a direct density interpretation while making the method computationally efficient for large samples. We study the score under contamination and show that regular positive-density observations and contamination-driven observations have distinct asymptotic orders. Simulations over a broad range of contamination models illustrate these theoretical regimes, show competitive performance relative to standard benchmarks, and document computing time. A credit-card fraud application with 29 variables illustrates that the method works well on a large real data set.

Figures

Figures reproduced from arXiv: 2607.14335 by Aur\'elien Nicosia, Michel Carbon, Thierry Duchesne.

Figure 1
Figure 1. Figure 1: Distribution of Di by class, sample size, inlier model, and λ. Rows correspond to data￾generating models and columns to sample sizes. Small λ corresponds to weak local support among outliers; large λ corresponds to stronger mutual support. The y-axis is on a logarithmic scale. 3.2 Robustness to contamination We next compare LOO-LBFP with standard baselines under four contamination scenarios. The goal is no… view at source ↗
Figure 2
Figure 2. Figure 2: Representative realizations of the four simulated scenarios: (a) Mixture, three Gaus [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Execution time as a function of sample size. Curves are method-specific, and both axes [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

13 extracted references · 3 canonical work pages

  1. [1]

    Bolton, R. J. & Hand, D. J. (2002), ‘Statistical fraud detection: A review’,Statistical Science 17(3), 235–255. 24

    2002

  2. [2]

    M., Kriegel, H., Ng, R

    Breunig, M. M., Kriegel, H., Ng, R. T. & Sander, J. (2000), ‘Lof: Identifying density-based local outliers’,ACM SIGMOD Record29(2), 93–104

    2000

  3. [3]

    & Duchesne, T

    Carbon, M. & Duchesne, T. (2024), ‘Multivariate frequency polygon for stationary random fields’, Annals of the Institute of Statistical Mathematics76(2), 263–287. URL:https://doi.org/10.1007/s10463-023-00883-5

    work page doi:10.1007/s10463-023-00883-5 2024

  4. [4]

    & Duchesne, T

    Carbon, M. & Duchesne, T. (2025), ‘Asymptotic normality of multivariate frequency polygons for stationary random fields’,Annals of the Institute of Statistical Mathematics. URL:https://link.springer.com/article/10.1007/s10463-025-00952-x

    work page doi:10.1007/s10463-025-00952-x 2025

  5. [5]

    & Kumar, V

    Chandola, V., Banerjee, A. & Kumar, V. (2009), ‘Anomaly detection: A survey’,ACM computing surveys (CSUR)41(3), 1–58. Dal Pozzolo, A., Caelen, O., Johnson, R. A. & Bontempi, G. (2015), ‘Calibrating probability with undersampling for unbalanced classification’,2015 IEEE Symposium Series on Computational Intelligencepp. 159–166. European cardholders dataset...

    2009

  6. [6]

    Hampel, F. R. (1974), ‘The influence curve and its role in robust estimation’,Journal of the American Statistical Association69(346), 383–393. URL:https://doi.org/10.1080/01621459.1974.10482962

    arXiv 1974

  7. [7]

    Huber, P. J. (1964), ‘Robust estimation of a location parameter’,The Annals of Mathematical Statistics35(1), 73–101. 25

    1964

  8. [8]

    & Hyndman, R

    Kandanaarachchi, S. & Hyndman, R. J. (2021), ‘Leave-one-out kernel density estimates for outlier detection’,Journal of Computational and Graphical Statistics31(2), 586–599

    2021

  9. [9]

    T., Ting, K

    Liu, F. T., Ting, K. M. & Zhou, Z.-H. (2008), Isolation forest,in‘2008 Eighth IEEE International Conference on Data Mining’, IEEE, pp. 413–422. URL:https://doi.org/10.1109/ICDM.2008.17

    doi:10.1109/icdm.2008.17 2008

  10. [10]

    & Shim, K

    Ramaswamy, S., Rastogi, R. & Shim, K. (2000), Efficient algorithms for mining outliers from large data sets,in‘Proceedings of the 2000 ACM SIGMOD International Conference on Management of Data’, ACM, New York, NY, USA, pp. 427–438

    2000

  11. [11]

    & Kriegel, H

    Schubert, E., Zimek, A. & Kriegel, H. (2014), Generalized outlier detection with flexible kernel density estimates (kdeos),in‘Proceedings of the 2014 SIAM International Conference on Data Mining (SDM)’, pp. 542–550. URL:https://epubs.siam.org/doi/10.1137/1.9781611973440.63

    work page doi:10.1137/1.9781611973440.63 2014

  12. [12]

    Scott, D. W. (1985), ‘Frequency polygons, theory and applications’,Journal of the American Statistical Association80, 348–354

    1985

  13. [13]

    N., Bischl, B

    Vanschoren, J., van Rijn, J. N., Bischl, B. & Torgo, L. (2013), ‘OpenML: Networked science in machine learning’,SIGKDD Explorations15(2), 49–60. URL:https://doi.org/10.1145/2641190.2641198 26

    arXiv 2013

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.