Online robust covariance matrix estimation and outlier detection
Paul Guillot, Antoine Godichon-Baggioni, Stéphane Robin, Laure Sansonnet · 2026
Read Online robust covariance matrix estimation and outlier detection by Paul Guillot, Antoine Godichon-Baggioni, Stéphane Robin, Laure Sansonnet (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-03957.txt.
Introduction
Paul Guillot and three co-authors present, to their knowledge, the first online method for robust covariance estimation and outlier detection on streaming data. Covariance matrices summarize how variables move together, and outliers can badly bias them. Worse, outliers mask each other, hiding in the very distortion they cause. Existing robust methods were designed for fixed datasets processed all at once, not for points arriving one by one.
The new method estimates the geometric median and variance simultaneously as data streams in. Each incoming point gets a Mahalanobis distance, a scale-aware measure of strangeness, before being judged an outlier or not. Robust estimates of center and spread are what keep the masking effect in check during that live decision. Streaming robust statistics serve fraud monitoring, sensor networks, and any system that cannot wait to clean its data.
Catalog record
| CREATOR | Paul Guillot, Antoine Godichon-Baggioni, Stéphane Robin, Laure Sansonnet |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 51 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-03957
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Online robust covariance matrix estimation and outlier detection Paul Guillot, Antoine Godichon-Baggioni, Stéphane Robin, Laure Sansonnet Abstract Robust estimation of the covariance matrix and detection of outliers remain major challenges in statistical data analysis, particularly when the proportion of contaminated observations increases with the size of the dataset. Outliers can severely bias parameter estimates and induce a masking effect , whereby some outliers conceal the presence of other outliers, further complicating their detection. Although many approaches have been proposed for covariance estimation and outlier detection, to our knowledge, none of these methods have been implemented in an online setting. In this paper, we focus on online covariance matrix estimation and outlier detection. Specifically, we propose a new method for simultaneously and online estimating the geometric median and variance, which allows us to calculate the Mahalanobis distance for each incoming data point before deciding whether it should be considered an outlier. To mitigate the masking effect, robust estimation techniques for the mean and variance are required. Our approach uses the geometric median for robust estimation of the location and the median covariance matrix for robust estimation of the dispersion parameters. The new online methods proposed for parameter estimation and outlier detection allow real-time identification of outliers as data are observed sequentially. The performance of our methods is demonstrated on simulated datasets. 1 Sorbonne Université, Université Paris Cité, CNRS, Laboratoire de Probabilités, Statistique et Modélisation, LPSM, F-75005 Paris, France 2 CREST, CNRS, ENSAE, IP Paris, Palaiseau, France 1 Introduction Aim and framework. In the multivariate Gaussian setting, the covariance structure plays a central role in describing data dependence and variability. Classical estimators such as the sample mean and covariance matrix are highly sensitive to outliers, motivating the need for robust alternatives. Indeed, the acquisition of large-scale data in high-dimensional spaces is unfortunately often accompanied by contamination. The automatic detection of these atypical data is not simple, and the use of robust techniques is an interesting alternative, in particular for online outlier detection. The paper objective is then to adapt and extend robust statistical procedures to the online setting, ensuring both computational efficiency and resistance to contamination. Robust estimation of the covariance matrix. Two main approaches can be distinguished for robust covariance estimation. The first family consists of modifications of the empirical covariance matrix to improve its robustness. This includes, for example, the Minimum Covariance Determinant ( Rousseeuw, 1985 , MCD:) or shrinkage approaches ( Ledoit and Wolf, 2004 ) . The second family replaces classical covariance and variance estimates with robust alternatives, such as the comedian and the median absolute deviation ( Falk, 1997 ) , improvements yielding positive definite estimates ( Cabana et al., 2021 ; Maronna and Zamar, 2002 , see) . To the best of our knowledge, none of these methods currently admit an online implementation. The approach we propose is based one the geometric median and the median covariation matrix. The geometric median is a robust measure of location ( Haldane, 1948 ; Kemperman, 1987 ) that can be preferred to the mean because it has a 50 \% breakdown point ( Gervini, 2006 ) . It has been extensively studied, and many methods have been proposed to estimate it, both in offline settings Weiszfeld, 1937 ; Vardi and Zhang, 2000 ; Beck and Sabach, 2015 and online settings Cardot et al., 2013 ; Godichon-Baggioni and Lu, 2023 . The median covariance matrix, in turn, was introduced by Kraus and Panaretos, 2012 , and both offline and online estimation procedures were later proposed by Cardot and Godichon-Baggioni, 2017 . More recently, Godichon-Baggioni and Robin, 2022 used it to reconstruct a robust offline estimator of the variance. Outlier detection. A first family of outlier detection methods relies on dimensionality reduction techniques, assuming that outliers are primarily concentrated along certain principal components ( Friedman and Tukey, 1974 ; Stahel, 1981 ; Donoho, 1982 ; Caussinus and Ruiz, 1990 ; Tyler et al., 2009 ; Peña and Prieto, 2001 ; Hubert et al., 2005 ) . These approaches are largely distribution-free, but may be computationally demanding, and therefore challenging to implement in an online context. In contrast, we focus here on Mahalanobis distance–based approaches, which account for the covariance structure of the data ( Jolliffe, 1986 ) . In this setting, an observation is flagged as an outlier when its associated Mahalanobis distance exceeds a certain threshold. To mitigate the masking effect, robust estimates of location and scatter are required ( Rousseeuw and Van Zomeren, 1990 ) . A key advantage of this approach is that it can be efficiently implemented in an online framework. Contribution: a novel online approach. In this work, we propose an algorithm to estimate all the quantities of interest (the median, the median covariation matrix, the variance-covariance matrix) in an online manner, while simultaneously enabling online outlier detection. Although this method is highly accurate, its computational cost can be high if an orthonormalization step is required at each iteration. To address this issue, we introduce a streaming version of our method, which processes data arriving sequentially in batches. This reduces the reliance on costly orthonormalization steps and allows the overall complexity to be reduced to the usual \mathcal{O}(nd^{2}) complexity, where n is the number of observation and d is the data dimension. Our method also enables real-time detection of outliers in incoming data without the need to recompute previous estimates from scratch. Outline. The paper is organized as follows. A review of existing offline procedures, including the geometric median and the median covariation matrix, is presented in Section 2 . The online and the streaming versions of the novel algorithm are described in Section 3 . Finally, Section 4 is devoted to numerical experiments assessing the performances and the accuracy of the proposed method. 2 Classical offline methods In this section, we provide more details about the principal robust approaches for covariance matrix estimation and outlier detection, with particular attention to their existing formulations in an online framework. 2.1 Robust estimation of the covariance matrix Sample covariance based methods. The sample mean and covariance matrix are highly sensitive to outliers, motivating robust alternatives based on modified covariance estimators. The minimum covariance determinant (MCD) estimator ( Rousseeuw, 1985 ) , in the context of elliptical distributions, seeks the subset of h observations with the smallest covariance determinant, yielding robust estimates of multivariate location and scatter. However, its exact computation is combinatorial, and the FAST-MCD algorithm ( Rousseeuw and Driessen, 1999 ) was proposed to approximate it efficiently via iterative C-steps. As far as we know, no online version currently exists. Other approaches, in the context of functional datas, include the trimmed covariance estimator ( Gervini, 2012 ) , which excludes detected outliers, and the shrinkage estimator ( Ledoit and Wolf, 2004 ) , which combines the sample covariance with a structured target to improve robustness. As far as we know, no online version of these methods has been proposed. Comedian-based methods. Another family of methods for estimating location and scatter parameters relies on the Comedian approach, which replaces classical covariance and variance with robust alternatives: the Comedian ( \operatorname{COM} ) and the squared Median Absolute Deviation ( \operatorname{MAD}^{2} ). For random variables U and V , these