pith. sign in

arxiv: 2606.29524 · v1 · pith:ZUJGLUC5new · submitted 2026-06-28 · 📊 stat.ME

Modelling and detecting mild and gross anomalies in circular data via double-contaminated models

Antonio Punzo , Andri\"ette Bekker , Arno Otto , Priyanka Nagar , Cristina Tortora This is my paper

Pith reviewed 2026-06-30 01:55 UTC · model grok-4.3

classification 📊 stat.ME
keywords circular datamixture modelsanomaly detectionrobust statisticsdirectional datacontamination modelsvon Miseswrapped normal
0
0 comments X

The pith

A three-component mixture model jointly captures mild and gross anomalies in circular data for improved fit and decomposition.

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

This paper develops a model-based approach to handle both mild and gross anomalies in circular data by extending standard unimodal distributions into mixtures. The framework includes a gross-anomaly model with a uniform component, a mild-anomaly model with a less concentrated component sharing the mean direction, and a combined double-contaminated model. These allow automatic classification of observations using posterior probabilities and provide mixing weights as measures of anomaly prevalence. Evaluated on simulations and real datasets of animal movements and wind directions, the joint modelling enhances fit and reveals latent structure. Readers would care because it reframes robustness as a tool for interpretation rather than mere outlier removal in directional statistics.

Core claim

The central discovery is that finite mixture models constructed from a reference circular distribution by adding a uniform component for gross anomalies and a less concentrated version sharing the same mean for mild anomalies, including their combination in a double-contaminated model, enable robust inference where posterior component probabilities classify observations and mixing weights quantify anomaly levels, as shown in simulations and applications to directional data.

What carries the argument

The double-contaminated model: a three-component finite mixture of a reference distribution, its less concentrated counterpart with identical mean direction, and a circular uniform distribution.

If this is right

  • Posterior probabilities provide automatic classification of observations as reference, mild anomaly, or gross anomaly without needing thresholds.
  • Mixing weights serve as interpretable measures of the prevalence of each type of anomaly and the inflation of dispersion.
  • Jointly modelling both anomaly types yields better model fit compared to modelling only one type.
  • The approach identifies latent structure in the directional data beyond simple anomaly detection.

Where Pith is reading between the lines

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

  • This framework could be extended to other circular distributions or to linear data with analogous contamination structures.
  • The decomposition into components might aid in identifying behavioral patterns in movement data that correspond to different anomaly types.
  • If the shared mean direction assumption is relaxed, more flexible models could capture directional shifts in anomalies.

Load-bearing premise

The reference distribution is unimodal and symmetric, mild anomalies are a less concentrated version with exactly the same mean direction, and gross anomalies follow a uniform distribution.

What would settle it

A simulation where data generated with mild anomalies having a different mean direction shows the double-contaminated model fails to improve fit or misclassifies observations would falsify the utility of the construction.

Figures

Figures reproduced from arXiv: 2606.29524 by Andri\"ette Bekker, Antonio Punzo, Arno Otto, Cristina Tortora, Priyanka Nagar.

Figure 1
Figure 1. Figure 1: Probability density functions of WN-based mixture models for anomaly detection. The reference WN distribution [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Anomaly detection map for n = 100 and κ = 100. (a) Example of simulated WN data, with red points indicating the 32 equally spaced locations at which additional points may be placed on the circle. (b) Detection heatmap where each cell corresponds to the modal classification (across 100 replications) for two added points located at positions (θ1, θ2). Green: no anomalies detected (WN); orange: gross anomaly … view at source ↗
Figure 3
Figure 3. Figure 3: Anomaly detection map for n = 100 and κ = 10. (a) Example of simulated WN data, with red points indicating the 32 equally spaced locations at which additional points may be placed on the circle. (b) Detection heatmap where each cell corresponds to the modal classification (across 100 replications) for two added points located at positions (θ1, θ2). Green: no anomalies detected (WN); orange: gross anomaly o… view at source ↗
Figure 4
Figure 4. Figure 4: Mean squared error (MSE) for estimating µ (top row) and κ (bottom row) with n = 100 and κ = 100. The left column displays results under the WN model, while the right column corresponds to the cuWN model. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Mean squared error (MSE) for estimating µ (top row) and κ (bottom row) with n = 100 and κ = 10. The left column displays results under the WN model, while the right column corresponds to the cuWN model. Overall, these results justify the necessity of the proposed framework, demonstrating that a model capable of handling both mild and gross anomalies simultaneously is essential. Results for a larger sample … view at source ↗
Figure 6
Figure 6. Figure 6: Anomaly detection results for the Marion Island dataset under the uWN model. Observations assigned to the reference [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Anomaly detection results for the larva dataset under the uWN, cWN, and cuWN models. Observations assigned to [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Anomaly detection results for the Col de la Roa wind dataset under the uWN, cWN, and cuWN models. Circular plots [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
read the original abstract

In this paper, we propose a model-based framework to robustify inference for circular data in the presence of anomalous observations, distinguishing between mild and gross anomalies. Starting from a unimodal and symmetric reference model on $[0,2\pi)$, parametrized by a mean direction and concentration, we construct a family of finite mixtures: a gross-anomaly model obtained by adding a circular uniform component; a mild-anomaly (contaminated) model obtained by mixing the reference distribution with a less concentrated version sharing the same mean direction; and a general three-component specification combining both models, the double-contaminated model. Posterior component probabilities provide an automatic classification of observations without ad hoc thresholds, while the mixing weights yield interpretable measures of anomaly prevalence and dispersion inflation. For illustration, we consider two classical circular reference distributions, the wrapped normal and von Mises. The methodology is evaluated through an extensive simulation study and three real-data applications involving animal movement directions and wind directions. The results indicate that jointly modelling mild and gross departures improves model fit and yields an informative decomposition of the directional data, demonstrating that mixture-based robustness is valuable not only for anomaly detection but also for the interpretation and the identification of latent structure in directional data.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes a model-based framework for robust inference on circular data by distinguishing mild and gross anomalies via finite mixture models. Starting from a unimodal symmetric reference distribution (von Mises or wrapped normal) parametrized by mean direction and concentration, it constructs a gross-anomaly model by adding a uniform component, a mild-anomaly model by mixing with a less concentrated version sharing the same mean, and a three-component double-contaminated model combining both. Posterior component probabilities enable automatic classification, while mixing weights provide measures of anomaly prevalence; the approach is illustrated on simulations and three real datasets (animal movement and wind directions), with the claim that joint modelling improves fit and yields an informative decomposition of latent structure.

Significance. If the central claims hold after addressing identifiability, the work would provide a principled extension of mixture-based robustness to directional data, enabling both anomaly detection and interpretable decomposition into mild versus gross departures. This could be valuable for applications in ecology and environmental science where distinguishing anomaly types aids interpretation beyond mere outlier removal.

major comments (2)
  1. [§2 (model construction)] §2 (model construction): the three-component double-contaminated model is defined by mixing the reference f(·; μ, κ), a mild component f(·; μ, κ') with κ' < κ, and a uniform gross component. When the fitted κ' approaches zero the mild and gross components become nearly indistinguishable, risking label-switching and unstable posterior probabilities. This directly threatens the abstract's claim that the model 'yields an informative decomposition' of mild versus gross anomalies, as opposed to simply improving in-sample fit via extra parameters. The manuscript must either impose identifiability constraints (e.g., κ' ≥ δ > 0) or demonstrate via targeted simulations that posterior classification remains reliable when κ' is small.
  2. [Simulation study and real-data sections] Simulation study and real-data sections: the abstract states that 'jointly modelling mild and gross departures improves model fit and yields an informative decomposition,' yet the provided text supplies no numerical results, recovery rates for mixing weights, or comparisons of posterior component stability across scenarios where κ' is near zero. Without such evidence the support for the decomposition claim cannot be verified and the weakest assumption (that mild anomalies are adequately captured by a less concentrated component sharing exactly the same mean) remains untested under boundary conditions.
minor comments (2)
  1. [Abstract] The abstract would benefit from a brief statement of the key numerical findings (e.g., improvement in log-likelihood or classification accuracy) from the simulation study and the three real-data examples.
  2. [Model specification] Notation for the concentration parameters (κ vs. κ') should be introduced with an explicit statement of the ordering constraint κ' < κ to avoid ambiguity in the mild-anomaly component definition.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and constructive report. The two major comments raise important issues regarding identifiability and empirical support for the decomposition claims. We address each point below and outline the revisions we will make.

read point-by-point responses

  1. Referee: [§2 (model construction)] §2 (model construction): the three-component double-contaminated model is defined by mixing the reference f(·; μ, κ), a mild component f(·; μ, κ') with κ' < κ, and a uniform gross component. When the fitted κ' approaches zero the mild and gross components become nearly indistinguishable, risking label-switching and unstable posterior probabilities. This directly threatens the abstract's claim that the model 'yields an informative decomposition' of mild versus gross anomalies, as opposed to simply improving in-sample fit via extra parameters. The manuscript must either impose identifiability constraints (e.g., κ' ≥ δ > 0) or demonstrate via targeted simulations that posterior classification remains reliable when κ' is small.

    Authors: We agree that the risk of label-switching when κ' approaches zero is a substantive concern that could weaken the interpretability of the mild-versus-gross decomposition. In the revised manuscript we will impose a lower bound κ' ≥ 0.1 during estimation (implemented via constrained optimization) and will add a brief discussion of this constraint in Section 2. We will also include a small set of targeted simulations that deliberately set the true κ' close to the boundary and report posterior classification accuracy and mixing-weight recovery under those conditions. revision: yes

  2. Referee: [Simulation study and real-data sections] Simulation study and real-data sections: the abstract states that 'jointly modelling mild and gross departures improves model fit and yields an informative decomposition,' yet the provided text supplies no numerical results, recovery rates for mixing weights, or comparisons of posterior component stability across scenarios where κ' is near zero. Without such evidence the support for the decomposition claim cannot be verified and the weakest assumption (that mild anomalies are adequately captured by a less concentrated component sharing exactly the same mean) remains untested under boundary conditions.

    Authors: The existing simulation study reports overall model-fit improvements and mixing-weight recovery across a range of contamination levels, but we acknowledge that it does not contain the specific boundary-case experiments requested (κ' near zero, posterior stability metrics). We will add these targeted simulation results, including recovery rates and classification stability tables, to the revised simulation section. The real-data analyses already contain posterior-probability summaries; we will augment them with the same stability diagnostics where relevant. revision: yes

Circularity Check

0 steps flagged

Explicit mixture construction; no load-bearing reduction to fitted inputs

full rationale

The paper defines its double-contaminated model by direct construction: a three-component mixture of the reference distribution, a less-concentrated version sharing the mean direction, and a uniform component. Posterior probabilities and mixing weights are then computed from the fitted model as outputs. No equation equates a claimed result (e.g., anomaly decomposition) back to the mixing weights by definition, no self-citation supplies a uniqueness theorem, and no fitted parameter is relabeled as an independent prediction. The central claim therefore rests on the explicit parametric form rather than on circular re-use of its own fitted values.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The framework rests on the choice of reference distribution (wrapped normal or von Mises) and the specific forms chosen for the anomaly components; these are domain modeling decisions rather than derived quantities.

free parameters (2)
  • mixing weights
    Proportions of the gross-anomaly, mild-anomaly, and reference components are estimated from data.
  • concentration parameters
    Concentration of the reference distribution and the less-concentrated mild-anomaly component are fitted.
axioms (1)
  • domain assumption Reference distribution is unimodal and symmetric on [0,2π)
    Explicitly stated as the starting point for constructing all mixture variants.

pith-pipeline@v0.9.1-grok · 5756 in / 1332 out tokens · 34076 ms · 2026-06-30T01:55:42.496078+00:00 · methodology

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

61 extracted references

  1. [1]

    Econometrics and Statistics , volume=

    A tractable, parsimonious and flexible model for cylindrical data, with applications , author=. Econometrics and Statistics , volume=. 2017 , publisher=

    2017

  2. [2]

    Statistical Papers , volume=

    Sine-skewed circular distributions , author=. Statistical Papers , volume=. 2011 , publisher=

    2011

  3. [3]

    Biometrika , volume=

    A tractable and interpretable four-parameter family of unimodal distributions on the circle , author=. Biometrika , volume=. 2015 , publisher=

    2015

  4. [4]

    Bernoulli , pages=

    An extended family of circular distributions related to wrapped Cauchy distributions via Brownian motion , author=. Bernoulli , pages=. 2013 , publisher=

    2013

  5. [5]

    A family of distributions on the circle with links to, and applications arising from, M

    Kato, Shogo and Jones, MC , journal=. A family of distributions on the circle with links to, and applications arising from, M. 2010 , publisher=

    2010

  6. [6]

    Ecological Indicators , volume=

    Circular distributions of fallen logs as an indicator of forest disturbance regimes , author=. Ecological Indicators , volume=. 2012 , publisher=

    2012

  7. [7]

    2017 , publisher=

    Modern Directional Statistics , author=. 2017 , publisher=

    2017

  8. [8]

    and Neuh

    Pewsey, A. and Neuh. Circular Statistics in. 2013 , publisher=

    2013

  9. [9]

    2013 , publisher=

    Outlier Analysis , author=. 2013 , publisher=

    2013

  10. [10]

    2015 , volume=

    Robust Cluster Analysis and Variable Selection , author=. 2015 , volume=

    2015

  11. [11]

    Journal of the American Statistical Association , volume=

    The identification of multiple outliers , author=. Journal of the American Statistical Association , volume=. 1993 , publisher=

    1993

  12. [12]

    2013 , publisher=

    Identification of Outliers , author=. 2013 , publisher=

    2013

  13. [13]

    Journal of Applied Statistics , volume=

    Dichotomous unimodal compound models: application to the distribution of insurance losses , author=. Journal of Applied Statistics , volume=. 2020 , publisher=

    2020

  14. [14]

    2011 , publisher=

    Robust Statistics , author=. 2011 , publisher=

    2011

  15. [15]

    Journal of Classification , year =

    Melnykov, Yana , title =. Journal of Classification , year =

  16. [16]

    , title =

    Lim, Heeju and Ordoñez, José Alejandro and Punzo, Antonio and Lachos, Victor H. , title =. Journal of Computational and Graphical Statistics , year =

  17. [17]

    Fr\"uhwirth-Schnatter, Sylvia , title =

  18. [18]

    Journal of Classification , year =

    Zhang, Yingying and Melnykov, Volodymyr and Melnykov, Igor , title =. Journal of Classification , year =

  19. [19]

    Statistics and Probability Letters , year =

    Dong, Aqi and Melnykov, Volodymyr , title =. Statistics and Probability Letters , year =

  20. [20]

    2009 , publisher=

    Directional Statistics , author=. 2009 , publisher=

    2009

  21. [21]

    Metrika , volume=

    The large-sample joint distribution of key circular statistics , author=. Metrika , volume=. 2004 , publisher=

    2004

  22. [22]

    Mixture models, outliers, and the

    Aitkin, Murray and Wilson, Granville Tunnicliffe , journal=. Mixture models, outliers, and the. 1980 , publisher=

    1980

  23. [23]

    Journal of classification , volume=

    Fixed point clusters for linear regression: computation and comparison , author=. Journal of classification , volume=. 2002 , publisher=

    2002

  24. [24]

    1994 , publisher=

    Outliers in statistical data , author=. 1994 , publisher=

    1994

  25. [25]

    Journal of the American Statistical Association , volume=

    The masking breakdown point of multivariate outlier identification rules , author=. Journal of the American Statistical Association , volume=. 1999 , publisher=

    1999

  26. [26]

    Statistics in Transition , volume=

    Clustering methods: From classical models to new approaches , author=. Statistics in Transition , volume=

  27. [27]

    Gallegos, Mar. Trimmed. Sankhy. 2009 , publisher=

    2009

  28. [28]

    2014 , publisher=

    Robust cluster analysis and variable selection , author=. 2014 , publisher=

    2014

  29. [29]

    Trimmed k -means:

    Cuesta-Albertos, Juan Antonio and Gordaliza, Alfonso and Matr. Trimmed k -means:. The Annals of Statistics , volume=. 1997 , publisher=

    1997

  30. [30]

    2005 , journal=

    A robust method for cluster analysis , author=. 2005 , journal=

    2005

  31. [31]

    Journal of Computational and Graphical Statistics , volume=

    Trimming tools in exploratory data analysis , author=. Journal of Computational and Graphical Statistics , volume=. 2003 , publisher=

    2003

  32. [32]

    The Annals of Statistics , volume=

    A general trimming approach to robust cluster Analysis , author=. The Annals of Statistics , volume=

  33. [33]

    Advances in Data Analysis and Classification , volume=

    A review of robust clustering methods , author=. Advances in Data Analysis and Classification , volume=. 2010 , publisher=

    2010

  34. [34]

    Advances in Data Analysis and Classification , volume=

    The influence function of the TCLUST robust clustering procedure , author=. Advances in Data Analysis and Classification , volume=. 2012 , publisher=

    2012

  35. [35]

    Journal of the Royal Statistical Society: Series C (Applied Statistics) , volume=

    Outliers in circular data , author=. Journal of the Royal Statistical Society: Series C (Applied Statistics) , volume=. 1980 , publisher=

    1980

  36. [36]

    Robust estimation of the concentration parameter of the von

    Ko, Daijin , journal=. Robust estimation of the concentration parameter of the von. 1992 , publisher=

    1992

  37. [37]

    Book of the Short Papers-SIS 2023 Ancona , pages=

    Anomaly detection in Circular Data , author=. Book of the Short Papers-SIS 2023 Ancona , pages=. 2023 , publisher=

    2023

  38. [38]

    Communications in Statistics-Simulation and Computation , volume=

    Statistics for a new test of discordance in circular data , author=. Communications in Statistics-Simulation and Computation , volume=. 2012 , publisher=

    2012

  39. [39]

    Journal of Modern Applied Statistical Methods , volume=

    Detection of outliers in univariate circular data using robust circular distance , author=. Journal of Modern Applied Statistical Methods , volume=

  40. [40]

    Communications in Statistics-Simulation and Computation , volume=

    A new test of discordancy in circular data , author=. Communications in Statistics-Simulation and Computation , volume=. 2009 , publisher=

    2009

  41. [41]

    Communications in Statistics-Simulation and Computation , volume=

    A new discordancy test in circular data using spacings theory , author=. Communications in Statistics-Simulation and Computation , volume=. 2016 , publisher=

    2016

  42. [42]

    Repository CRAN , volume=

    Package ‘circular’ , author=. Repository CRAN , volume=

  43. [43]

    2001 , publisher=

    Topics in Circular Statistics , author=. 2001 , publisher=

    2001

  44. [44]

    Computational Statistics & Data Analysis , volume=

    Robust estimation for circular data , author=. Computational Statistics & Data Analysis , volume=. 2007 , publisher=

    2007

  45. [45]

    Jones, MC and Pewsey, Arthur , journal=. Inverse. 2012 , publisher=

    2012

  46. [46]

    The Annals of Statistics , pages=

    Identifiability of finite mixtures for directional data , author=. The Annals of Statistics , pages=. 1983 , publisher=

    1983

  47. [47]

    Identifiability of finite mixtures-with applications to circular distributions , author=. Sankhy. 2004 , publisher=

    2004

  48. [48]

    R: A language and environment for statistical computing

  49. [49]

    Choosing starting values for the

    Biernacki, Christophe and Celeux, Gilles and Govaert, G. Choosing starting values for the. Computational Statistics & Data Analysis , volume=. 2003 , publisher=

    2003

  50. [50]

    Biometrical Journal , volume=

    Parsimonious mixtures of multivariate contaminated normal distributions , author=. Biometrical Journal , volume=. 2016 , publisher=

    2016

  51. [51]

    Environmetrics , volume=

    Modeling Bounded Count Environmental Data Using a Contaminated Beta-Binomial Regression Model , author=. Environmetrics , volume=. 2026 , publisher=

    2026

  52. [52]

    1974 , publisher=

    A new look at the statistical model identification , author=. 1974 , publisher=

    1974

  53. [53]

    Estimating the dimension of a model , author=. The. 1978 , publisher=

    1978

  54. [54]

    Model-based

    Banfield, Jeffrey D and Raftery, Adrian E , journal=. Model-based. 1993 , publisher=

    1993

  55. [55]

    Journal of Machine Learning Research , volume=

    Consistency, breakdown robustness, and algorithms for robust improper maximum likelihood clustering , author=. Journal of Machine Learning Research , volume=

  56. [56]

    Estimation and computations for

    Coretto, Pietro , journal=. Estimation and computations for. 2022 , publisher=

    2022

  57. [57]

    Journal of Classification , volume=

    On model-based clustering of directional data with heavy tails , author=. Journal of Classification , volume=. 2023 , publisher=

    2023

  58. [58]

    Robust clustering in regression analysis via the contaminated

    Punzo, Antonio and McNicholas, Paul D , journal=. Robust clustering in regression analysis via the contaminated. 2017 , publisher=

    2017

  59. [59]

    Statistical Methods in Medical Research , volume=

    A contaminated regression model for count health data , author=. Statistical Methods in Medical Research , volume=. 2025 , publisher=

    2025

  60. [60]

    Statistics & Probability Letters , volume=

    Contaminated Kent mixture model for clustering non-spherical directional data with heavy tails or scatter , author=. Statistics & Probability Letters , volume=. 2024 , publisher=

    2024

  61. [61]

    Ecological Modelling , volume=

    Investigation of ecologically relevant wind patterns on Marion Island using Computational Fluid Dynamics and measured data , author=. Ecological Modelling , volume=. 2022 , publisher=

    2022