Pith. sign in

REVIEW 3 major objections 4 minor 61 references

seMCD: Sequentially implemented Monte Carlo depth computation with statistical guarantees

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Depth computation can be done sequentially, returning a bucket that contains the true value with probability at least 1−α at a fraction of the Monte Carlo cost.

desk verdict Useful depth-computation method with exact Bernoulli-case guarantees, but the general-case guarantee is imported from a self-cited theorem and deserves a careful referee check. read the letter →

arxiv 2507.06227 v1 pith:4TX2G7ZP submitted 2025-07-08 stat.ME math.STstat.COstat.MLstat.TH

classification stat.MEmath.STstat.COstat.MLstat.TH MSC 62L1062G0562H12
keywords SequentiallyimplementedMonteCarloDepth-bucketsAnytime-validprobabilisticguaranteesHigh-dimensionaldataOutlierdetectionAnomalyClassificationStatisticaldepth
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that expensive depth computations can be replaced by a sequential Monte Carlo routine that stops when enough evidence has accumulated, returning a preselected interval, a seMCD-bucket, known to contain the true depth value with probability at least $1-\alpha$. For depth functions whose Monte Carlo summands are Bernoulli indicators, such as simplicial, spherical, lens, and band depths, the guarantee is exact and finite-sample; for more general E-depth functions with finite second moments, including integrated depths, the modified band depth, and classification contrasts, the guarantee is asymptotic. The practical claim is that these guarantees arrive at a fraction of the usual Monte Carlo cost: stopping times in the experiments are typically in the hundreds rather than the standard $10^5$ or $100d$ fixed budgets. A curious reader should care because depth-based outlier detection, classification, and depth-region computation currently either use huge Monte Carlo budgets or lack any uncertainty statement for the resulting decisions.

What carries the argument

The engine is the seMCD-process $S_N = \sum_{j=1}^N H(\xi_j)$ together with boundary sequences $L_N^{(h)}$ and $U_N^{(h)}$ that encircle $Nh$; the algorithm stops when $S_N$ leaves the corridor between the boundaries associated with adjacent split points. A monotonicity condition on the boundaries, namely that higher split points give higher upper and lower boundaries, transfers the elementary two-bucket guarantee to arbitrary multi-bucket configurations. In the Bernoulli case the boundaries are computed recursively from an error-spending sequence $\alpha_N$ so that each side's first-crossing probability is kept below $\alpha_N/2$. In the nonparametric case the boundaries are $L^{(h)}_{N,m}(\tilde{\sigma}_N^2) = Nh - \tilde{\sigma}_N w_{(\gamma_1,\gamma_2)}(N,m)$ and $U^{(h)}_{N,m}(\tilde{\sigma}_N^2) = Nh + \tilde{\sigma}_N w_{(\gamma_1,\gamma_2)}(N,m)$, with $w_{(\gamma_1,\gamma_2)}(N,m) = c_\alpha(\gamma_1,\gamma_2) m^{\gamma_2-1/2} N^{\gamma_1}(m+N)^{1-\gamma_1-\gamma_2}$, where $c_\alpha$ is the $(1-\alpha)$-quantile of $\sup_{0 \le t \le 1} |B(t)|/(t^{\gamma_1}(1-t)^{\gamma_2})$ for a standard Brownian bridge; these boundaries satisfy the monotonicity condition and yield the asymptotic uniform coverage.

What would settle it

For the Bernoulli case, set the true $h$ exactly at a split point, run the non-overlapping algorithm repeatedly with $\alpha=0.05$, and check that the stopping frequency does not exceed $0.05$; for the asymptotic case, choose a depth whose summands $H(\xi)$ have infinite variance, run with $m=500$, and check whether false-bucket rates stay below $\alpha$.

Watch

Extended reading notes

Core claim

The central claim is that many depth computations, theoretical or empirical, multivariate or functional, fit the template $h = E(H(\xi))$, and that this template can be evaluated by a sequential Monte Carlo procedure that stops as soon as a partial-sum process $S_N = \sum_{j=1}^N H(\xi_j)$ crosses precomputed boundary sequences. The output is not a point estimate but a user-specified interval, the seMCD-bucket, and the guarantee is attached to the bucket: $P_h(h \in B) \ge 1-\alpha$. For Type A depths with an indicator kernel, $H(\xi)$ is Bernoulli and the boundary sequences are built recursively from an error-spending sequence, giving an exact finite-sample guarantee. For all other E-depths with finite second moments, boundary sequences of the form $N h \pm \tilde{\sigma}_N c_\alpha(\gamma_1,\gamma_2) m^{\gamma_2-1/2} N^{\gamma_1}(m+N)^{1-\gamma_1-\gamma_2}$ deliver the same coverage asymptotically. The empirical sections show bucket decisions made with far fewer Monte Carlo draws than standard fixed budgets, while false-bucket rates stay below the nominal tolerance.

Load-bearing premise

For the general nonparametric version of the method, the $1-\alpha$ guarantee is inherited from a theorem in a previous paper that is assumed but not reproved here; that theorem requires the Monte Carlo summands to have finite second moments and a non-degenerate variance, and the guarantee only holds as the burn-in parameter grows.

Editorial extensions

If this is right

  • For every Type A depth with an indicator kernel, a user who fixes tolerance $\alpha$ and a family of buckets obtains a bucket guaranteed to contain the true depth with probability at least $1-\alpha$ in finite samples.
  • For general E-depths with finite second moments, including integrated depths, the modified band depth, and classification differences, the same coverage holds asymptotically as the burn-in parameter $m$ grows.
  • In the reported experiments, typical stopping times are far below the standard Monte Carlo budgets of $10^5$ or $100d$, with runtimes often one to two orders of magnitude smaller.
  • For anomaly detection and maximum-depth classification, outputting a bucket instead of an exact depth value is sufficient for the decision, and the bucket decision carries the $\alpha$ error guarantee.
  • The point estimate at the stopping time is biased and should not be interpreted as an unbiased Monte Carlo mean; only the output bucket carries the statistical guarantee.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: the only structural requirement is that the target be an expectation, so the bucket machinery transfers to any Monte Carlo estimand beyond depth; the depth setting supplies natural decision buckets, but the sequential-testing core is generic.
  • Editorial extension: for a fixed burn-in $m=500$, the asymptotic guarantee is a limit statement, so users with a specific depth should calibrate $m$ and the spending parameters by simulation before relying on nominal coverage.
  • Editorial extension: overlapping buckets eliminate the non-stopping problem at split points, suggesting a general design principle: replace point estimates with a user-chosen family of overlapping intervals and let sequential sampling decide among them at controlled error.
  • Editorial extension: because simplicial depth contracts as dimension grows, the algorithm inherits the curse of dimensionality; a practical next step is an adaptive rescaling rule for split points rather than only the fixed $1/2^{d-1}$ scaling mentioned in the paper.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes seMCD, a sequential Monte Carlo method for computing statistical depth functions and depth-based quantities that outputs a user-specified bucket containing the target value with probability at least 1−α. For Type A depth functions with indicator kernel, where the summands are Bernoulli, the algorithm has an exact finite-sample guarantee via recursively constructed boundary sequences. For general E-depth functions with finite second moments, the paper imports an asymptotic time-uniform boundary from Gnettner and Kirch (2025) and derives an asymptotic coverage guarantee. The method is demonstrated on modified band depth, integrated rank-weighted depth, anomaly detection on the KDD smtp data, and binary classification. The central claims are that seMCD requires fewer Monte Carlo samples than plain-vanilla methods while providing probabilistic guarantees on the output bucket.

Significance. If the asymptotic guarantee is valid, this is a useful contribution to depth computation: it gives a practical, computationally cheaper alternative to plain-vanilla Monte Carlo with a user-controlled error probability, and it covers a broad class of E-depth functions. The Bernoulli-case construction is careful and provides exact guarantees. The paper ships reproducible code and reports false-bucket rates in simulations, which is a strength. The main weakness is that the non-parametric half of the guarantee rests on an unproved, self-cited theorem whose conditions are not verified in the present manuscript.

major comments (3)
  1. [Section 4.2, Theorem 4.4 and Eq. (20)] The asymptotic coverage claim is the load-bearing step for all non-Bernoulli applications (modified band depth, IRW depth, classification differences), but Theorem 4.4 is not proved here. The text says only that Gnettner and Kirch (2025, Theorem 3.1) "yields after some elementary algebra" the boundary sequences in (20). Because that source is a self-citation and the specific plug-in-variance form is not derived, the conditions under which the per-N estimator \tilde\sigma_N^2 can be used in place of the true variance at every N are not verified. Please either prove Theorem 4.4 in an appendix or restate the source theorem in full and verify its assumptions for the depth summands in Sections 2.1–2.2 and for the online variance estimator.
  2. [Section 3.2.5, Greedy seMCD-algorithm] The simulations in Section 5 all use the greedy algorithm, yet the claim that Lemma 3.2's coverage guarantee holds for it is only asserted, not proved: the paper states "The assertion and proof of Lemma 3.2 and thus the corresponding statistical guarantee also hold for the greedy algorithm" without giving the argument. Because the greedy updates restrict to the current feasible interval (l,r), a short proof or a formal reduction to Lemma 3.2 is needed to make the empirical claims load-bearing.
  3. [Section 4.2 and Section 5.1] The boundary sequences (20) are justified only in the limit m→∞, but the recommended value m=500 is used throughout the simulations without a sensitivity analysis. Please report the empirical false-bucket-rate for a range of m (e.g., 100, 500, 1000) to show that the approximation is adequate at the recommended value. Without this, the finite-sample performance in Tables 1–3 is an undocumented empirical assertion.
minor comments (4)
  1. [Section 4.1.2, Lemma 4.2] The displayed inequality has the upper and lower boundaries reversed: it should read L^{(h)}_\ell < S_\ell < U^{(h)}_\ell, not U^{(h)}_\ell < S_\ell < L^{(h)}_\ell.
  2. [Remark 3.4] The sentence "this is no longer the case for d ⩾ 1" should refer to d ⩾ 5, consistent with the contraction rate O(1/2^d) and the preceding sentence.
  3. [Equation (13)] The displayed overlapping buckets are garbled in the typesetting; please replace with a clear list of intervals, e.g., (−∞, h2), (h1, h3), (h2, h4), ... .
  4. [Section 5.2] The notation "5−4" and "15−4" is ambiguous; use 5×10^{-4} and 15×10^{-4} or equivalent.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular step: the coverage guarantees are derived by induction on recursively built boundaries, and the cited self-authored theorem is applied as independent published support.

full rationale

The central coverage derivation is not circular. In the Bernoulli case, Section 4.1 constructs boundary sequences by an explicit recursion on p_N(k), β_U(N), and β_L(N), and Lemma 4.2 verifies condition (I) by induction; (17) then controls each misclassification direction, with Lemma 3.2 and Corollary 4.3 extending the bound to multi-bucket and overlapping outputs. In the general E-depth case, Theorem 4.4 is stated by citing Gnettner and Kirch (2025, Theorem 3.1) rather than reproved; although this is a self-citation, it is a published theorem with explicit assumptions (iid summands and finite non-zero variance) that do not include the target coverage, so it functions as independent mathematical support rather than a circular premise. The choices γ1=0.1, γ2=0.4, m=500, and lm=10 are user-specified inputs used before evaluating false-bucket rates; no reported FBR is fed back to redefine the boundaries. Consequently, no equation defines the output bucket to be its own guarantee and no fitted parameter is renamed as a prediction. Concerns about the plug-in variance step in Theorem 4.4 are potential correctness risks, not circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method's guarantees rest on the representation of depth quantities as expectations, the validity of the sequential-testing boundary sequences, and, for the non-parametric case, an imported theorem from the authors' own prior work. The user-chosen tuning parameters listed above affect stopping time rather than theoretical coverage, but they are hand-picked in the simulations rather than derived.

free parameters (3)
  • Error spending sequence parameter kappa = 1000 (default)
    Spending sequence alpha_N = alpha * N / (kappa + N) in eq. (11) shapes the Bernoulli-case boundaries; kappa is a user choice and 1000 is used in simulations.
  • Boundary shape parameters gamma1 and gamma2 = 0.1 and 0.4 (simulations)
    Non-parametric boundaries (20) depend on gamma1 and gamma2, controlling early and late conservativeness. The asymptotic coverage holds for 0 <= gamma1, gamma2 < 1/2, but stopping time and false-bucket behavior depend on the choice.
  • Burn-in parameter m and minimal variance sample l_m = m = 500, l_m = 10 (simulations)
    The asymptotic statement (I*) takes m to infinity; in practice m = 500 and l_m = 10 are chosen per Remark 4.5. These affect finite-sample behavior but not the asymptotic guarantee.
assumptions (5)
  • domain assumption Gnettner and Kirch (2025, Theorem 3.1) is valid as applied, including the online variance plug-in, yielding the time-uniform crossing probability (I*).
    This theorem is the entire basis for the non-parametric guarantees in Theorem 4.4 and is not proved in this paper; it is first-author self-citation.
  • domain assumption The quantity of interest can be written as E(H(xi)) or F(E(H(xi))) for known computable H and samplable xi.
    Section 2 defines E-depth and monotonic E-depth functions as the scope of the method; depth functions outside this class, e.g. Tukey depth, are explicitly excluded in Section 3.1.
  • domain assumption For finite-sample guarantees, H(xi) has a Bernoulli distribution with unknown success probability h.
    Section 4.1 requires Bernoulli summands for the exact recursive boundary construction; this covers Type A depths with indicator kernel such as simplicial, spherical, lens and beta-skeleton depth.
  • domain assumption For non-parametric guarantees, H(xi_j) are iid with 0 < Var(H(xi_1)) < infinity.
    Theorem 4.4 requires iid summands with finite non-degenerate variance; the paper states this is typically fulfilled in the depth context.
  • standard math The boundary sequences satisfy the monotonicity condition Mon or Mon(h1,...,hk), required by Lemma 3.2.
    Non-parametric boundaries (20) clearly satisfy Mon by their form; Bernoulli boundaries satisfy Mon(h1,...,hk) for alpha <= 1/4, with an adjustment for early N mentioned but not detailed in Corollary 4.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of seMCD: Sequentially implemented Monte Carlo depth computation with statistical guarantees." pith.science (2026). https://pith.science/paper/4TX2G7ZP

@misc{pith2026250706227,
  author       = {Pith},
  title        = {Pith review of: seMCD: Sequentially implemented Monte Carlo depth computation with statistical guarantees},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4TX2G7ZP}},
  note         = {Machine review of arXiv:2507.06227}
}
read the original abstract

Statistical depth functions provide center-outward orderings in spaces of dimension larger than one, where a natural ordering does not exist. The numerical evaluation of such depth functions can be computationally prohibitive, even for relatively low dimensions. We present a novel sequentially implemented Monte Carlo methodology for the computation of, theoretical and empirical, depth functions and related quantities (seMCD), that outputs an interval, a so-called seMCD-bucket, to which the quantity of interest belongs with a high probability prespecified by the user. For specific classes of depth functions, we adapt algorithms from sequential testing, providing finite-sample guarantees. For depth functions dependent on unknown distributions, we offer asymptotic guarantees using non-parametric statistical methods. In contrast to plain-vanilla Monte Carlo methodology the number of samples required in the algorithm is random but typically much smaller than standard choices suggested in the literature. The seMCD method can be applied to various depth functions, covering multivariate and functional spaces. We demonstrate the efficiency and reliability of our approach through empirical studies, highlighting its applicability in outlier or anomaly detection, classification, and depth region computation. In conclusion, the seMCD-algorithm can achieve accurate depth approximations with few Monte Carlo samples while maintaining rigorous statistical guarantees.

Figures

Figures reproduced from arXiv: 2507.06227 by the authors.

Figure 1
Figure 1. Upper boundary sequence {U (0.25) N : N ∈ N} (blue dashed-dotted) and lower boundary sequence {L (0.25) N : N ∈ N} (blue dashed) for tolerance α = 0.01 and H(ξ) ∼ Bernoulli(0.25). Three seMCD-sample paths {SN : N ∈ N} representing different outcomes for stopping the algorithm: orange sequences. with, for example, h1 = 0.05; and (2) binary maximum depth classification, with h1 = 0 corresponding to equal depth within … view at source ↗
Figure 2
Figure 2. Stopping regions (white) for non-overlapping seMCD-buckets (left panel) and overlapping [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Boxplots for the number of samples at time of stopping for the modified band depth. [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Boxplots for the point estimates hˆ τ for the modified band depth. The dash-dotted line in the boxplots indicated the exact value. The plot shows that these values have to be interpreted with care. that, we complement this information in [PITH_FULL_IMAGE:figures/full_…
Figure 5
Figure 5. Figure 5: Boxplots for the number of samples taken at time of stopping for the seMCD-algorithms for [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: Three-dimensional scatter plots of the smtp data set. [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]
Figure 7
Figure 7. Figure 7: Boxplots for each considered feature in the banana data set. [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 42 canonical work pages

  1. [1]

    URL https://kdd.ics.uci.edu/databases/kddcup99/kddcup99.html

    Kdd C up 1999 D ata, 1999. URL https://kdd.ics.uci.edu/databases/kddcup99/kddcup99.html. Retrieved April 29, 2025

  2. [2]

    Afshani, D

    P. Afshani, D. R. Sheehy, and Y. Stein. Approximating the simplicial depth in high dimensions. In The European Workshop on Computational Geometry, 2016

  3. [3]

    Aloupis, C

    G. Aloupis, C. Cortés, F. Gómez, M. Soss, and G. Toussaint. Lower bounds for computing statistical depth. Computational Statistics & Data Analysis, 40 0 (2): 0 223--229, 2002. ISSN 0167-9473. doi:10.1016/S0167-9473(02)00032-4

  4. [4]

    Arribas-Gil and J

    A. Arribas-Gil and J. Romo. Shape outlier detection and visualization for functional data: the outliergram . Biostatistics, 15 0 (4): 0 603--619, 03 2014. ISSN 1465-4644. doi:10.1093/biostatistics/kxu006

  5. [5]

    Aue and C

    A. Aue and C. Kirch. The state of cumulative sum sequential changepoint testing 70 years after Page . Biometrika, page asad079, 12 2023. ISSN 1464-3510. doi:10.1093/biomet/asad079

  6. [6]

    Baharav and T

    T. Baharav and T. L. Lai. Adaptive data depth via multi-armed bandits. Journal of Machine Learning Research, 24 0 (155): 0 1--29, 2023

  7. [7]

    Briend, G

    S. Briend, G. Lugosi, and R. I. Oliveira. On the quality of randomized approximations of T ukey's depth. arXiv preprint arXiv:2309.05657v2, 2023. doi:10.48550/arXiv.2309.05657

  8. [8]

    Cascos and I

    I. Cascos and I. Molchanov. Multivariate risks and depth-trimmed regions. Finance and Stochastics, 11 0 (3): 0 373--397, Jul 2007. ISSN 1432-1122. doi:10.1007/s00780-007-0043-7

Show all 61 references
  1. [9]

    Chakraborty and P

    A. Chakraborty and P. Chaudhuri. The spatial distribution in infinite dimensional spaces and related quantiles and depths . The Annals of Statistics, 42 0 (3): 0 1203 -- 1231, 2014. doi:10.1214/14-AOS1226

  2. [10]

    Y. Chen, X. Dang, H. Peng, and H. Bart. Outlier detection with the kernelized spatial depth function. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31 0 (2): 0 288--305, 2009. doi:10.1109/TPAMI.2008.72

  3. [11]

    A. Y. Cheng and M. Ouyang. On algorithms for simplicial depth. In Proceedings of the 13th Canadian Conference on Computational Geometry (CCCG'01), pages 53--56, 2001

  4. [12]

    C.-S. J. Chu, M. Stinchcombe, and H. White. Monitoring structural change. Econometrica, 64 0 (5): 0 1045--1065, 1996. ISSN 00129682, 14680262. doi:10.2307/2171955

  5. [13]

    Cl \'e men c on, P

    S. Cl \'e men c on, P. Mozharovskyi, and G. Staerman. Affine invariant integrated rank-weighted statistical depth: properties and finite sample analysis . Electronic Journal of Statistics, 17 0 (2): 0 3854 -- 3892, 2023. doi:10.1214/23-EJS2189

  6. [14]

    Cuesta-Albertos and A

    J. Cuesta-Albertos and A. Nieto-Reyes. The random T ukey depth. Computational Statistics & Data Analysis, 52 0 (11): 0 4979--4988, 2008. ISSN 0167-9473. doi:10.1016/j.csda.2008.04.021

  7. [15]

    Cuevas and R

    A. Cuevas and R. Fraiman. On depth measures and dual statistics. a methodology for dealing with general data. Journal of Multivariate Analysis, 100 0 (4): 0 753--766, 2009. ISSN 0047-259X. doi:10.1016/j.jmva.2008.08.002

  8. [16]

    Cuevas, M

    A. Cuevas, M. Febrero, and R. Fraiman. Robust estimation and classification for functional data via projection-based depth notions. Computational Statistics, 22 0 (3): 0 481--496, Sep 2007. ISSN 1613-9658. doi:10.1007/s00180-007-0053-0

  9. [17]

    Dang and R

    X. Dang and R. Serfling. Nonparametric depth-based multivariate outlier identifiers, and masking robustness properties. Journal of Statistical Planning and Inference, 140 0 (1): 0 198--213, 2010. ISSN 0378-3758. doi:10.1016/j.jspi.2009.07.004

  10. [18]

    D. Ding, A. Gandy, and G. Hahn. A simple method for implementing monte carlo tests. Computational Statistics, 35 0 (3): 0 1373--1392, Sep 2020. ISSN 1613-9658. doi:10.1007/s00180-019-00927-6

  11. [19]

    Dyckerhoff and P

    R. Dyckerhoff and P. Mozharovskyi. Exact computation of the halfspace depth. Computational Statistics & Data Analysis, 98: 0 19--30, 2016. ISSN 0167-9473. doi:10.1016/j.csda.2015.12.011

  12. [20]

    R. T. Elmore, T. P. Hettmansperger, and F. Xuan. Spherical data depth and a multivariate median. In Data Depth: Robust Multivariate Analysis, Computational Geometry and Applications, volume 72 of DIMACS Series in Discrete Mathematics and Theoretical Computer Science, pages 87-...

  13. [21]

    Febrero, P

    M. Febrero, P. Galeano, and W. González-Manteiga. Outlier detection in functional data by depth measures, with application to identify abnormal nox levels. Environmetrics, 19 0 (4): 0 331--345, 2008. doi:10.1002/env.878

  14. [22]

    Fischer and A

    L. Fischer and A. Ramdas. Sequential M onte C arlo testing by betting. Journal of the Royal Statistical Society Series B: Statistical Methodology, page qkaf014, 04 2025. ISSN 1369-7412. doi:10.1093/jrsssb/qkaf014

  15. [23]

    Franke, M

    J. Franke, M. Hefter, A. Herzwurm, K. Ritter, and S. Schwaar. Adaptive quantile computation for B rownian bridge in change-point analysis. Computational Statistics & Data Analysis, 167: 0 107375, 2022. ISSN 0167-9473. doi:10.1016/j.csda.2021.107375

  16. [24]

    A. Gandy. Sequential implementation of M onte C arlo tests with uniformly bounded resampling risk. Journal of the American Statistical Association, 104 0 (488): 0 1504--1511, 2009. doi:10.1198/jasa.2009.tm08368

  17. [25]

    Gandy, G

    A. Gandy, G. Hahn, and D. Ding. Implementing M onte C arlo tests with p-value buckets. Scandinavian Journal of Statistics, 47 0 (3): 0 950--967, 2020. doi:10.1111/sjos.12434

  18. [26]

    A. K. Ghosh and P. Chaudhuri. On maximum depth and related classifiers. Scandinavian Journal of Statistics, 32 0 (2): 0 327--350, 2005. doi:10.1111/j.1467-9469.2005.00423.x

  19. [27]

    Gnettner

    F. Gnettner. Depth functions for multivariate and functional data: computation and statistical inference. PhD thesis, Otto-von-Guericke-Universität Magdeburg, 2024

  20. [28]

    Gnettner and C

    F. Gnettner and C. Kirch. A new and flexible class of sharp asymptotic time-uniform confidence sequences. Statistics & Probability Letters, 226: 0 110462, 2025. ISSN 0167-7152. doi:10.1016/j.spl.2025.110462

  21. [29]

    González-De La Fuente , A

    L. González-De La Fuente , A. Nieto-Reyes, and P. Terán. Statistical depth for fuzzy sets. Fuzzy Sets and Systems, 443: 0 58--86, 2022. ISSN 0165-0114. doi:10.1016/j.fss.2021.09.015. Fuzzy Intervals and Applications

  22. [30]

    Horváth, M

    L. Horváth, M. Hušková, P. Kokoszka, and J. Steinebach. Monitoring changes in linear models. Journal of Statistical Planning and Inference, 126 0 (1): 0 225--251, 2004. ISSN 0378-3758. doi:10.1016/j.jspi.2003.07.014

  23. [31]

    Hubert, P

    M. Hubert, P. J. Rousseeuw, and P. Segaert. Multivariate functional outlier detection. Statistical Methods & Applications , 24 0 (2): 0 177--202, Jul 2015. ISSN 1613-981X. doi:10.1007/s10260-015-0297-8

  24. [32]

    Jörnsten

    R. Jörnsten. Clustering and classification based on the L_1 data depth. Journal of Multivariate Analysis, 90 0 (1): 0 67--89, 2004. ISSN 0047-259X. doi:10.1016/j.jmva.2004.02.013. Special Issue on Multivariate Methods in Genomic Data Analysis

  25. [33]

    Banana quality: Can you identify good bananas by their numerical characteristics?, 2024

    Kaggle.com. Banana quality: Can you identify good bananas by their numerical characteristics?, 2024. URL https://www.kaggle.com/datasets/l3llff/banana. Uploaded by L3LLFF, retrieved April 5, 2024

  26. [34]

    T. L. Lai. On Confidence Sequences . The Annals of Statistics, 4 0 (2): 0 265 -- 280, 1976. doi:10.1214/aos/1176343406

  27. [35]

    K. K. G. Lan and D. L. DeMets. Discrete sequential boundaries for clinical trials. Biometrika, 70 0 (3): 0 659--663, 1983. ISSN 00063444. doi:10.2307/2336502

  28. [36]

    Lange, K

    T. Lange, K. Mosler, and P. Mozharovskyi. Fast nonparametric classification based on data depth. Statistical Papers, 55 0 (1): 0 49--69, Feb 2014. ISSN 1613-9798. doi:10.1007/s00362-012-0488-4

  29. [37]

    J. Li, J. A. Cuesta-Albertos, and R. Y. Liu. DD -classifier: Nonparametric classification procedure based on DD -plot. Journal of the American Statistical Association, 107 0 (498): 0 737--753, 2012. doi:10.1080/01621459.2012.688462

  30. [38]

    F. T. Liu, K. M. Ting, and Z.-H. Zhou. Isolation forest. In 2008 E ighth IEEE I nternational C onference on D ata M ining , pages 413--422. IEEE, 2008

  31. [39]

    R. Y. Liu. On a Notion of Data Depth Based on Random Simplices . The Annals of Statistics, 18 0 (1): 0 405 -- 414, 1990. doi:10.1214/aos/1176347507

  32. [40]

    R. Y. Liu and K. Singh. A Q uality I ndex B ased on D ata D epth and M ultivariate R ank T ests. Journal of the American Statistical Association, 88 0 (421): 0 252--260, 1993. ISSN 01621459. doi:10.2307/2290720

  33. [41]

    Liu and R

    Z. Liu and R. Modarres. Lens data depth and median. Journal of Nonparametric Statistics, 23 0 (4): 0 1063--1074, 2011. doi:10.1080/10485252.2011.584621

  34. [42]

    L \'o pez-Pintado and J

    S. L \'o pez-Pintado and J. Romo. On the concept of depth for functional data. Journal of the American Statistical Association, 104 0 (486): 0 718--734, 2009. ISSN 01621459. doi:10.1198/jasa.2009.0108

  35. [43]

    Marsaglia

    G. Marsaglia. Choosing a Point from the Surface of a Sphere . The Annals of Mathematical Statistics, 43 0 (2): 0 645 -- 646, 1972. doi:10.1214/aoms/1177692644

  36. [44]

    K. Mosler. Depth statistics. In C. Becker, R. Fried, and S. Kuhnt, editors, Robustness and Complex Data Structures: Festschrift in Honour of Ursula Gather, pages 17--34, Berlin, Heidelberg, 2013. Springer Berlin Heidelberg. ISBN 978-3-642-35494-6. doi:10.1007/978-3-642-35494-6_2

  37. [45]

    Mozharovskyi and R

    P. Mozharovskyi and R. Valla. Anomaly detection using data depth: multivariate case. International Journal of Data Science and Analytics, May 2025. ISSN 2364-4168. doi:10.1007/s41060-025-00784-1

  38. [46]

    Nieto-Reyes and H

    A. Nieto-Reyes and H. Battey. A T opologically V alid D efinition of D epth for F unctional D ata. Statistical Science, 31 0 (1): 0 61--79, 02 2016. doi:10.1214/15-STS532

  39. [47]

    Nieto-Reyes and H

    A. Nieto-Reyes and H. Battey. A topologically valid construction of depth for functional data. Journal of Multivariate Analysis, 184: 0 104738, 2021. ISSN 0047-259X. doi:10.1016/j.jmva.2021.104738

  40. [48]

    Nieto-Reyes and J

    A. Nieto-Reyes and J. Cabrera. Statistical depth based normalization and outlier detection of gene expression data. arXiv preprint arXiv:2206.13928, 2022. doi:10.48550/arXiv.2206.13928

  41. [49]

    H. Oja. Descriptive statistics for multivariate distributions. Statistics & Probability Letters, 1 0 (6): 0 327--332, 1983. ISSN 0167-7152. doi:/10.1016/0167-7152(83)90054-8

  42. [50]

    Ramsay, S

    K. Ramsay, S. Durocher, and A. Leblanc. Integrated rank-weighted depth. Journal of Multivariate Analysis, 173: 0 51--69, 2019. ISSN 0047-259X. doi:10.1016/j.jmva.2019.02.001

  43. [51]

    H. Robbins. Statistical Methods Related to the Law of the Iterated Logarithm . The Annals of Mathematical Statistics, 41 0 (5): 0 1397 -- 1409, 1970. doi:10.1214/aoms/1177696786

  44. [52]

    Serfling

    R. Serfling. A depth function and a scale curve based on spatial quantiles. In Y. Dodge, editor, Statistical Data Analysis Based on the L1-Norm and Related Methods, pages 25--38, Basel, 2002. Birkh \"a user Basel. ISBN 978-3-0348-8201-9

  45. [53]

    Serfling and Y

    R. Serfling and Y. Zuo. General notions of statistical depth function . The Annals of Statistics, 28 0 (2): 0 461 -- 482, 2000. doi:10.1214/aos/1016218226

  46. [54]

    Sguera, P

    C. Sguera, P. Galeano, and R. Lillo. Spatial depth-based classification for functional data. TEST, 23 0 (4): 0 725--750, Dec 2014. ISSN 1863-8260. doi:10.1007/s11749-014-0379-1

  47. [55]

    S. C. Tan, K. M. Ting, and T. F. Liu. Fast anomaly detection for streaming data. In IJCAI proceedings-international joint conference on artificial intelligence , volume 22, page 1511. Citeseer, 2011

  48. [56]

    K. M. Ting, G.-T. Zhou, F. T. Liu, and J. S. C. Tan. Mass estimation and its applications. In Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 989--998, 2010

  49. [57]

    J. W. Tukey. Mathematics and the picturing of data. In R. D. James, editor, Proceedings of the International Congress of Mathematicians, volume 2, pages 523--531, Vancouver, 1974. Canadian Mathematical Congress

  50. [58]

    Williams, R

    G. Williams, R. Baxter, H. He, S. Hawkins, and L. Gu. A comparative study of rnn for outlier detection in data mining. In 2002 IEEE International Conference on Data Mining, 2002. Proceedings. , pages 709--712, 2002. doi:10.1109/ICDM.2002.1184035

  51. [59]

    Yamanishi, J.-i

    K. Yamanishi, J.-i. Takeuchi, G. Williams, and P. Milne. On-line unsupervised outlier detection using finite mixtures with discounting learning algorithms. Data Mining and Knowledge Discovery, 8 0 (3): 0 275--300, May 2004. ISSN 1573-756X. doi:10.1023/B:DAMI.0000023676.72185.7c

  52. [60]

    Yang and R

    M. Yang and R. Modarres. -skeleton depth functions and medians. Communications in Statistics-Theory and Methods, 47 0 (20): 0 5127--5143, 2018. doi:10.1080/03610926.2017.1386320

  53. [61]

    W. Zhao, Z. Xu, Y. Mu, Y. Yang, and W. Wu. Model-based statistical depth with applications to functional data. Journal of Nonparametric Statistics, 36 0 (2): 0 313--356, 2024. doi:10.1080/10485252.2023.2226262

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.