Exploring Multivariate Data Using Median Absolute Deviation Depth
summary
The gist
Exploring Multivariate Data Using Median Absolute Deviation Depth proposes the moving median absolute deviation (MMAD) as a robust depth construction based on the median absolute distance functional…
In short
The paper introduces Moving Median Absolute Deviation (MMAD) as a robust depth measure for multivariate data. It extends classical depth notions by using MMAD to define centrality, offering geometric insights into directional structure and providing a differentiable, scale-valued representation of centrality that is computationally efficient.
Key concepts
- Moving Median Absolute Deviation (MMAD)
- MMAD measures the dispersion of data around every point 'v' using the median instead of the mean. It identifies how much expansion is needed for 'v' to be a center, with smaller values indicating greater intrinsic centrality.
- Statistical Depth
- Statistical depth generalizes univariate order statistics to multivariate contexts, defining centrality through geometric probability constructs like Tukey’s halfspace depth. It provides a robust framework for classifying points based on their location within the data distribution.
- Directional Skewness
- This concept describes the imbalance of probability mass at the boundary of the central region defined by MMAD. Analyzing its directional derivatives reveals how probability mass is arranged along this boundary, offering insight into asymmetry in multivariate distributions.
Terminology used across episodes
This episode discusses
- Exploring Multivariate Data Using Median Absolute Deviation Depth · Paper Radio
- DepthProc An R Package for Robust Exploration of Multidimensional Economic Phenomena
The paper
Exploring Multivariate Data Using Median Absolute Deviation Depth · Read on arXiv
Elsayed A. H. Elamir
Department of Management and Marketing, College of Business Administration, Kingdom of Bahrain
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Exploring Multivariate Data Using Median Absolute Deviation Depth".
Jane: Here's the gist of the paper. Exploring Multivariate Data Using Median Absolute Deviation Depth proposes the moving median absolute deviation (MMAD) as a robust depth construction based on the median absolute distance functional with particular emphasis on its local geometry and probabilistic structure The MMAD depth identifies essentially the same central observations as classical depth notions while delivering additional information and geometric insight about directional structure The paper introduces the moving median absolute deviation (MMAD) functional, defined as a geometric descriptor of central region deformation employed to define a valid statistical depth It is shown that the MMAD remains fully compatible with axiomatic depth theory and extends the literature by introducing a differentiable, scale-valued representation of centrality The function measures how much expansion is required for a point to behave as a centre, where smaller values indicate intrinsic centrality Multivariate Extension and Geometry The framework extends from one dimension to multivariate data by defining the multivariate moving median absolute deviation (3MAD) functional, which measures how much the data must stretch for a point to act as a centre The geometric interpretation of 3MAD contours is determined by the chosen distance norm, such as L1, L2, or Mahalanobis distances The gradient of the functional provides a closed-form interpretation as the mean outward unit normal on the boundary of the 50% central region Validation and Comparison with Classical Depths Simulation studies empirically verify that 3MAD depth induces centre–outward orderings consistent with classical notions of statistical depth under idealized conditions The central region overlap analysis demonstrates that D3MAD preserves classical notions of centrality, achieving Jaccard overlaps at least 85% across all methods This confirms that D3MAD identifies essentially the same core observations as classical depth functions Directional Insights and Computational Efficiency The directional derivatives of the MMAD functional encode the imbalance of probability mass at the boundary of the 50% central region The spherical measure muv offers a distribution-valued description of how probability mass is arranged along the 3MAD boundary, revealing information about directional skewness From a computational perspective, 3MAD is attractive as it requires only distance calculations and a univariate median, resulting in an O(nd) complexity in the Euclidean case The distribution-valued depth is given by muv(A) = P(X - v X - v ∈ A X - v = Φ(v)) The study concludes that 3MAD positions itself as a depth measure that separates centrality from directional geometry, providing insight into how and where asymmetry arises in the distribution The spherical measure muv refines classical depth notions by furnishing a distribution-valued description of how probability mass is arranged along the 3MAD boundary The results position 3MAD as a depth measure that separates centrality from directional geometry The central region overlap analysis provides perhaps the most direct empirical validation of 3MAD as a depth measure Page 1 --- Exploring Multivariate Data Using Median Absolute Deviation Depth Elsayed A.
Tom: H.
Jane: Elamir Department of Management and Marketing, College of Business Administration, Kingdom of Bahrain Email: shabib@uob.edu.bh We propose and analyse the moving median absolute deviation (MMAD) as a robust depth construction based on the median absolute distance functional with particular emphasis on its local geometry and probabilistic structure In the univariate setting, we derive the derivative of the MMAD scale and interpret it through boundary mass imbalance, thereby establishing a direct connection to a robust skewness measure This idea extends naturally to multivariate setting that describes how observations are arranged along the 50% central region using directional derivative, a gradient representation and a spherical boundary distribution From a computational perspective, MMAD can be estimated efficiently using distance calculations without needing to complex optimization or projection schemes Multivariate applications based on depth correlations, contour visualizations and central region overlap demonstrate that MMAD identifies essentially the same central observations as classical depth notions while delivering additional information and geometric insight about directional structure These features make MMAD a practical and informative approach for robust multivariate data analysis Keywords: directional skewness; geometric data analysis; MAD; multivariate depth; robust nonparametric methods.
Tom: Subject classification code: 62G30 and 62G32 ORCID: 0000-0002-9430-072X Page 2 --- 1 Introduction Statistical depth was originally proposed as a generalization of univariate order statistics to multivariate contexts and has since developed into a broad and active field of both theoretical and applied research Early foundational work includes Tukey’s halfspace depth (Tukey, 1975) and Liu’s simplicial depth (Liu, 1990) that define centrality through geometric probability constructs These notions stimulated extensive research into depth-based inference, classification, and visualization, particularly due to their robustness and nonparametric nature (Nagy, 2025) (Zuo & Serfling, 2000) introduced a unifying axiomatic framework for statistical depth and formalizing fundamental properties such as affine invariance, maximality at the centre, monotonicity with respect to deepest points and vanishing at infinity This framework has become the standard benchmark for validating new depth notions and has guided the development of subsequent depth functions (Serfling, 2002) and (Shirke & Khorate, 2024) Among the most prominent outcomes of this line of work is projection depth which first motivated by (Donoho & Gasko, 1992), and later formalized by (Zuo, 2003) Projection depth characterizes centrality via worst-case standardized deviations along all directions and provides strong robustness guarantees (Jörnsten, 2004; Shirke & Khorate, 2023; Vencálek, 2017) Other important contributions include spatial depth and related notions based on spatial signs and spatial quantiles (Serfling, 2002; Vardi & Zhang, 2000) Since these depths rely on expectations rather than quantiles, this make them more sensitive to distributional shape and heavy tails although they are computationally attractive and yield smooth depth contours While the Mahalanobis depth is optimal under elliptical symmetry, it depends critically on covariance estimation Therefore, it is unstable in high-dimensional or contaminated settings (Liu et al., 1999; Zuo, 2021) and (Singh et al., 2022) Robust scale measures play a parallel role in multivariate analysis Median-based scale estimators are well known for their high breakdown point and bounded influence (Law et al., 1986; Rousseeuw & Croux, 1993) However, much of existing work considers these scale measures as global quantities computed at a fixed centre Recent surveys and monographs (Mosler, 2013; Zuo, 2003) emphasize that many statistical depth functions reduce complex geometric information to scalar rankings Although this is a desirable for ordering and classification, it restricts interpretability and diagnostic capability especially under skewness, multimodality, or cluster overlap Motivated by these considerations, we propose an approach built on the moving median absolute deviation (MMAD) with threefold contributions First, we introduce the MMAD functional as a geometric descriptor of central region deformation and employ it to define a valid statistical depth Second, we show that the MMAD remains fully compatible with axiomatic depth theory Third, we demonstrate how MMAD depth and its geometric diagnostics enable new tools for ranking, interpretable, and computationally tractable In this sense, MMAD depth complements rather than competes with classical depth measures It satisfies the axioms of statistical depth when normalized, aligns with projection depth under elliptical symmetry, and remains robust under heavy tails and contamination At the same time, it extends the literature by introducing a differentiable, scale-valued representation of centrality that enables explanation and diagnostics Page 3 --- 3 1 Introduction Statistical depth was originally proposed as a generalization of univariate order statistics to multivariate contexts and has since developed into a broad and active field of both theoretical and applied research Early foundational work includes Tukey’s halfspace depth (Tukey, 1975) and Liu’s simplicial depth (Liu, 1990) that define centrality through geometric probability constructs These notions stimulated extensive research into depth-based inference, classification, and visualization, particularly due to their robustness and nonparametric nature (Nagy, 2025) (Zuo & Serfling, 2000) introduced a unifying axiomatic framework for statistical depth and formalizing fundamental properties such as affine invariance, maximality at the centre, monotonicity with respect to deepest points and vanishing at infinity This framework has become the standard benchmark for validating new depth notions and has guided the development of subsequent depth functions (Serfling, 2002) and (Shirke & Khorate, 2024) Among the most prominent outcomes of this line of work is projection depth which first motivated by (Donoho & Gasko, 1992), and later formalized by (Zuo, 2003) Projection depth characterizes centrality via worst-case standardized deviations along all directions and provides strong robustness guarantees (Jörnsten, 2004; Shirke & Khorate, 2023; Vencálek, 2017) Other important contributions include spatial depth and related notions based on spatial signs and spatial quantiles (Serfling, 2002; Vardi & Zhang, 2000) Since these depths rely on expectations rather than quantiles, this make them more sensitive to distributional shape and heavy tails although they are computationally attractive and yield smooth depth contours While the Mahalanobis depth is optimal under elliptical symmetry, it depends critically on covariance estimation Therefore, it is unstable in high-dimensional or contaminated settings (Liu et al., 1999; Zuo, 2021) and (Singh et al., 2022) Robust scale measures play a parallel role in multivariate analysis Median-based scale estimators are well known for their high breakdown point and bounded influence (Law et al., 1986; Rousseeuw & Croux, 1993) However, much of existing work considers these scale measures as global quantities computed at a fixed centre Recent surveys and monographs (Mosler, 2013; Zuo, 2003) emphasize that many statistical depth functions reduce complex geometric information to scalar rankings Although this is a desirable for ordering and classification, it restricts interpretability and diagnostic capability especially under skewness, multimodality, or cluster overlap Motivated by these considerations, we propose an approach built on the moving median absolute deviation (MMAD) with threefold contributions First, we introduce the MMAD functional as a geometric descriptor of central region deformation and employ it to define a valid statistical depth Second, we show that the MMAD remains fully compatible with axiomatic depth theory Third, we demonstrate how MMAD depth and its geometric diagnostics enable new tools for ranking, interpretable, and computationally tractable In this sense, MMAD depth complements rather than competes with classical depth measures It satisfies the axioms of statistical depth when normalized, aligns with projection depth under elliptical symmetry, and remains robust under heavy tails and contamination At the same time, it extends the literature by introducing a differentiable, scale-valued representation of centrality that enables explanation and diagnostics Page 4 --- 4 2 Moving median absolute deviation (MMAD) Let X1, …, Xn be an independent and identically a random sample from a continuous distribution function Fx(.)(0 < F < 1), density fX(.)(f ≥ 0), quantile Q(F), mean μ = E(X), median Μ = QX(0.5), standard deviation σ = √E(x − μ)2, indicator function Ii≤k is 1 if i ≤ k, 0 else, and X(1), …, X(n) are the corresponding order statistics 2.1 Definition of MMAD For any candidate location v ∈ R, we define the moving median absolute deviation (MMAD) around v by GX(v) = Med(X − v), v ∈ R Unlike the classical median absolute deviation (MAD), which is evaluated only at the median of X, the function GX(v)is defined for every value v.
Paper summary: Tom: So we’ve gone through how the authors propose this moving median absolute deviation depth, and it’s clear they are building on existing concepts to create something that has both robustness and new directional information.
Jane: It seems like the title itself, Exploring Multivariate Data Using Median Absolute Deviation Depth, really captures the essence of what they're doing—it’s about using this median absolute deviation idea in a multivariate context.
Lu: The paper positions this method as complementing classical depth measures rather than competing with them because it satisfies the axioms of statistical depth when it is normalized and aligns with projection depth under elliptical symmetry.
Meng: I see what they mean by that, but I’m still wondering about the practical deployment at scale. If we are running a system where data streams in constantly, how stable is this moving median absolute deviation functional when the underlying distribution keeps shifting?
Tom: The authors emphasize its robustness under heavy tails and contamination, which suggests it handles those messy real-world data scenarios better than some of the more traditional methods they were comparing it to.
Jane: So what does this mean for us in terms of how we think about centrality in complex systems? It sounds like we get a better picture of the asymmetry in our data, not just where the average point is.
Lu: The paper suggests that this depth measure gives insight into precisely how and where asymmetry arises within a distribution.
Meng: If I had to deploy this right now, I’d want to know if the computational overhead stays low enough for high-throughput scenarios without needing massive pre-processing steps that add latency.
Tom: The paper does confirm it requires only distance calculations and a univariate median, which keeps the complexity manageable in the Euclidean case at O of n times d.
Jane: So we’re looking at a method that is mathematically sound according to depth theory, empirically validated by overlap comparisons, and computationally feasible without needing those heavy optimization schemes.
Lu: It extends the literature by introducing this differentiable, scale-valued representation of centrality which really enables better explanation and diagnostics for complex data structures.
Meng: So it’s a practical tool that offers more than just a rank; it offers a description of the shape of the central region in terms of directionality.
Tom: It seems like this paper gives us a concrete way to move from just finding where things are to understanding the geometry around them, which is valuable for any serious analysis.
Conclusion: Tom: So we’ve looked at how this paper uses median absolute deviation to build a depth measure for multivariate data, and now we need to talk about what it actually means.
Jane: It seems like they’re focusing on this moving median absolute deviation functional, which is really a new way to define centrality in data that isn't just looking at one single number.
Lu: Yeah, the authors are trying to take that basic idea of the median absolute distance and make it work for multiple dimensions.
Meng: I mean, they’re calling it a three-dimensional moving median absolute deviation functional when they move into the multivariate setting. That sounds complicated to implement on a production system.
Tom: It is complex, but the point is that this approach aims to be robust against those messy data distributions we see in the real world.
Jane: They show that this method stays compatible with existing theories of statistical depth and actually performs well in simulations compared to classical depths like projection depth.
Lu: The results show it preserves the same core observations as other methods, achieving jaccard overlaps of at least eighty five percent across different approaches they tested.
Meng: Eighty five percent overlap is good, but does that mean it’s a perfect deployment? What happens when the data distribution gets really weird?
Tom: They're saying it remains robust under heavy tails and contamination, which is a key part of what makes this functional interesting.
Jane: Beyond just finding the central points, they introduce something called a spherical measure, muv, to describe how that probability mass is arranged around the boundary of the central region.
Lu: That directional information from the spherical measure is what sets it apart—it gives us a geometric insight into where things are skewed.
Tom: So, basically, this paper positions this moving median absolute deviation as a tool that separates finding a central point from understanding how that data is shaped directionally.
Jane: Exactly. It’s not just about where the center is; it’s about the geometry around that center and what kind of asymmetry exists there.
Lu: It extends the literature by giving us this differentiable, scale-valued representation of centrality, which opens up new ways to diagnose data structures that classical methods miss.
Meng: I guess for a production system, having a way to visualize those directional insights instead of just getting one number is something we could use for debugging models.
Tom: It moves us from just ordering things to actually understanding the structure of the data itself, which is pretty significant for any serious analysis.
Jane: And that leads us right into how this directional skewness connects to other concepts in machine learning representation learning.
More episodes
- Daily Summary for 2026-11-19
- 2310.07973-Statistical Performance Guarantee for Subgroup Identification with Generic Machine Learning
- 2408.09155-Learning Robust Treatment Rules for Censored Data
- Daily Summary for 2026-11-19
- 2603.23397-Kinetic Langevin Splitting Schemes for Constrained Sampling
- 2606.09906-An information-geometric framework for mapping maximum potential biodiversity
- 2603.27718-Induced replication and the assessment of models
- 1307.0898-Entropy of a Zipfian Distributed Lexicon
- 2406.19974-Generalizing self-normalized importance sampling with couplings
- 2412.10213-Collaborative Optimal Design of Controlled Experiments with Subject Covariates