📊 statistics

Robust measures of dispersion for circular data with an anomaly detection rule

This paper extends three robust dispersion measures from linear to circular data to derive efficient estimators for von Mises and wrapped normal distributions, ultimately proposing a novel anomaly detection rule visualized via circular violin plots and validated through simulations and real-world datasets.

Original authors: Houyem Demni, Mia Hubert, Giovanni C. Porzio, Peter J. Rousseeuw

Published 2026-03-03

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Houyem Demni, Mia Hubert, Giovanni C. Porzio, Peter J. Rousseeuw

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are standing in the middle of a giant clock face. Instead of numbers, the clock has directions: North, South, East, West. Now, imagine you ask 20 people to point their fingers in the direction they think is "North." Most of them point roughly North, but a few are confused and point South, and one person is just spinning around pointing everywhere.

In statistics, this is called circular data. It's data that lives on a circle (like time of day, wind direction, or animal migration paths) rather than a straight line.

The problem? Standard math tools are terrible at handling these circles when a few people point the wrong way. If you calculate the "average" direction using normal math, those confused pointers can drag the average all the way to the opposite side of the clock, making the whole group look like they are pointing South when they are actually pointing North.

This paper introduces a new set of super-robust tools to measure how spread out these directions are, even when there are "bad apples" (outliers) in the mix. Here is the breakdown in simple terms:

1. The Problem: The "One Bad Apple" Effect

The authors start with a story about sea stars. Most of them swam toward the shore, but two swam in the opposite direction.

The Old Way (Circular Standard Deviation): If you use the old math, those two sea stars make the whole group look incredibly scattered. It's like if one person in a crowd of 100 standing still suddenly started running; the "average movement" of the crowd would look huge, even though 99 people were still.
The Goal: We need a way to measure the spread of the crowd that ignores the person running around.

2. The Solution: Three New "Robust" Rulers

The authors invented three new ways to measure spread on a circle. Think of them as three different types of rulers that refuse to be fooled by outliers.

CMAD (The "Middle" Ruler): This looks at the distance of every point from the median (the middle point) and picks the middle distance. It's like asking, "How far is the person in the middle of the crowd from the center?" It ignores the extreme runners.
CLMS (The "Tightest Half" Ruler): This is the star of the show. It asks: "What is the smallest slice of the circle that contains at least half of the people?" It effectively says, "Ignore the noisy edges; just look at the tightest group of 50%." It finds the core of the data and measures how tight that core is.
CLTS (The "Trimmed" Ruler): This calculates the spread of the tightest half of the data, similar to CLMS but uses a slightly different mathematical trick to find the "best" half.

The Winner: After testing them with simulations and real data, the CLMS (The "Tightest Half" Ruler) came out on top. It was the most accurate and the most resistant to bad data.

3. The New Tool: The "Circular Violin Plot"

Once you have a robust way to measure spread, you need a way to see the outliers. The authors created a new graph called a Circular Violin Plot.

What is a Violin Plot? Imagine a standard violin shape (wide in the middle, thin at the ends). In statistics, the width of the violin shows how many data points are in that area. A fat part means lots of data; a thin part means little data.
The Circular Twist: Instead of a straight violin, they wrapped it around the clock face.
- The thick part of the ring shows where most of the data is clustered (the "good" directions).
- The thin part shows where data is sparse.
- The Magic: They draw a "fence" (a cutoff line) around the thick part. Any data point (like a sea star) that falls outside this fence is automatically flagged as an outlier.

4. Real-World Results

The authors tested their new tools on three real datasets:

Northern Cricket Frogs: They found one frog that was trying to hop in the wrong direction. The new tool spotted it immediately.
Sardinian Sea Stars: This was the tricky one. Two sea stars were swimming the wrong way. Old methods missed one of them because the two "bad" stars hid each other (a problem called the "masking effect"). The new CLMS tool ignored both bad stars, calculated the true direction of the good ones, and flagged both outliers correctly.
Fruit Fly Larvae: They looked at how a fruit fly larva moved. The new plot showed that the data was too "messy" to fit a simple circle model. It suggested the fly was behaving in a more complex way than expected, helping scientists realize they needed a different mathematical model.

The Big Takeaway

This paper is about building unbreakable statistical tools for circular data. Just as you wouldn't use a flimsy ruler to measure a building under construction, you shouldn't use standard math to analyze directions when outliers are present.

By using the CLMS method and the Circular Violin Plot, scientists can now:

Find the true "center" of a group of directions.
Measure how tightly they are grouped without being scared by a few weird outliers.
Visually spot exactly who the "bad apples" are, even if there are two of them hiding together.

It's a bit like having a super-sensor that can see through the noise to find the true signal, no matter how much chaos is happening around it.

1. Problem Statement

Circular data (angles, directions, periodic observations) are prevalent in fields like biology and environmental sciences. Standard statistical measures for such data, specifically the Circular Standard Deviation (CSD), are highly sensitive to outliers. A single anomalous observation can drastically inflate the CSD, leading to misleading conclusions about the data's dispersion. While robust measures of location (e.g., the circular median) are well-established, robust measures of dispersion for circular data have received limited attention. Existing methods often lack theoretical robustness analysis (e.g., influence functions, breakdown values) or suffer from the "masking effect," where multiple outliers hide each other.

2. Methodology

The authors propose and analyze three robust dispersion measures adapted from linear statistics to the circular domain:

Circular Median Absolute Deviation (CMAD): Defined as the median of the shortest arc distances from the circular median.
Circular Least Median Spread (CLMS): An extension of the Least Median of Squares (LMS). It is defined as the minimum median distance from a point on the circle to the data, or equivalently, half the length of the shortest arc containing at least 50% of the probability mass.
Circular Least Trimmed Standard Deviation (CLTS): An extension of the Least Trimmed Squares (LTS). It is the minimum Circular Standard Deviation calculated over all arcs containing at least 50% of the data.

Theoretical Analysis:

Influence Functions (IF): The authors derived the IF for all three measures to quantify the effect of infinitesimal contamination. They found that CMAD and CLMS share the same IF structure (piecewise constant), while CLTS has a different shape (resembling $-\cos(\theta)$ near the mode).
Breakdown Values: They calculated the breakdown values (the maximum fraction of outliers a method can tolerate before failing).
- CMAD: Does not reach the theoretical maximum of 0.5 due to the requirement of a unique circular median.
- CLMS and CLTS: Both achieve the optimal breakdown value of 0.5, meaning they can tolerate up to 50% contamination.
Bias and Efficiency: The study computed relative bias curves under contamination and Asymptotic Relative Efficiency (ARE) compared to Maximum Likelihood Estimators (MLE).

Parameter Estimation & Anomaly Detection:

The authors constructed robust estimators for the concentration parameter ( $\kappa$ ) of the von Mises distribution and the dispersion parameter ( $\sigma$ ) of the Wrapped Normal distribution by inverting the relationships between the robust dispersion measures and the population CSD.
Anomaly Detection Rule: A rule was developed where a point is flagged as an outlier if its shortest arc distance from the robust circular median exceeds a cutoff value ( $c_\alpha$ ). This cutoff is derived from the 1- $\alpha/2$ quantile of the fitted robust model.
Visualization: A new graphical tool, the Circular Violin Plot, was introduced. It combines a non-parametric density estimate (mirrored radially) with a circular boxplot (interquartile range and median) to visualize data concentration and flag outliers.

3. Key Contributions

Extension of Robust Statistics: Successfully extended three classic robust scale estimators (MAD, LMS, LTS) to the circular domain and provided their first rigorous theoretical analysis (IF, bias, breakdown).
Identification of CLMS as Superior: Through simulation and theoretical comparison, the paper identifies CLMS as the best-performing estimator. It offers the highest breakdown value (0.5), competitive efficiency (approx. 37% at high concentration, similar to linear LMS), and the lowest relative bias under contamination.
Robust Anomaly Detection Framework: Developed a detection rule that avoids the masking effect common in sequential deletion methods. By relying on the CLMS (which uses only the central 50% of data), the method remains stable even with multiple outliers.
Circular Violin Plot: Introduced a novel visualization technique that effectively displays density, central tendency, and outliers on a circular manifold, overcoming the limitations of traditional circular boxplots.

4. Results

Simulation Studies:
- Under Mean Shift and Point Mass contamination models (10% and 20% outliers), the classical MLE was severely biased.
- The robust estimators (CMAD, CLMS, CLTS) remained stable.
- CLMS consistently outperformed CLTS and CMAD in terms of bias and stability, particularly in the Wrapped Normal and von Mises models.
Efficiency: The asymptotic relative efficiency of CLMS-based estimators approaches ~36.7% as concentration increases (matching the linear LMS efficiency), which is considered acceptable given the high robustness.
Real Data Applications:
1. Northern Cricket Frogs: Successfully identified a single known outlier using the CLMS-based cutoff, matching previous literature results but with a more robust framework.
2. Sardinian Sea Stars: The proposed method successfully identified two outliers that were missed by single-deletion techniques due to the masking effect. The CLMS-based cutoff was tight enough to flag both, whereas MLE-based cutoffs were inflated by the outliers.
3. Drosophila Fly Larva: The circular violin plot revealed that the data did not fit a von Mises distribution (heavy tails), suggesting a need for a different model (Inverse Batschelet), demonstrating the tool's utility in model checking.

5. Significance

This work fills a critical gap in directional statistics by providing a theoretically sound, robust framework for estimating dispersion and detecting outliers in circular data. The identification of CLMS as the optimal estimator provides practitioners with a reliable tool that is resistant to up to 50% contamination. Furthermore, the Circular Violin Plot offers a powerful visual diagnostic for circular data, enhancing the ability to detect anomalies and assess distributional assumptions in fields ranging from ecology to astronomy. The availability of R code ensures these methods are accessible for immediate application.