Multiple change-point detection on the circle via isolation using permutation testing

Sophia Loizidou, Andreas Anastasiou, Christophe LeyarXiv 2026 · DOI · arXiv

This paper introduces PCID (Permutation-based Circular Isolate-Detect), a new framework for identifying multiple "change-points"—sudden shifts in the mean—within circular or angular data. By combining a deterministic interval expansion strategy with non-parametric permutation testing, the authors provide a robust tool for analyzing periodic phenomena like wind directions, animal movement, or biological rhythms.

TL;DR

PCID is a novel algorithm designed to detect multiple structural shifts in circular data (angles) by first isolating potential change-points and then validating them via permutation testing. It overcomes the limitations of traditional linear methods and remains highly accurate even when the underlying noise distribution is unknown or serially correlated.

What Problem Does This Paper Address?

Most change-point detection algorithms are designed for data on the real line. However, many scientific fields deal with circular data—observations like compass headings, times of day, or phases of a cycle—where the values $0$ and $2\pi$ are identical.

Standard methods fail here because they don't respect this periodicity; for instance, a linear mean of $0.1$ and $6.2$ radians is roughly $3.15$, but the circular mean is nearly $0$. Furthermore, detecting multiple change-points is notoriously difficult because changes can "mask" one another (e.g., two opposite shifts might cancel out in a global test). While some circular methods exist, they often struggle with multiple changes or rely on strict assumptions about noise that don't hold in the real world.

Methods

The authors propose the PCID algorithm, which rests on three pillars:

  1. Isolation (Isolate-Detect): Instead of searching the entire dataset at once, PCID uses a deterministic "expanding window" strategy. It scans the data in slowly growing sub-intervals to ensure that, eventually, every change-point is isolated in at least one interval where no other change-points exist. This prevents the "masking" effect.
  2. Circular Contrast Function: The authors derived a specific contrast function (Equation 6 in the paper) based on the log-likelihood ratio of the von Mises distribution (the circular equivalent of the Normal distribution). This function uses circular sample means and resultant lengths to quantify the "strength" of a potential change.
  3. Permutation Testing: To decide if a detected change is statistically significant, PCID shuffles the data within the interval thousands of times. If the original contrast value is higher than almost all shuffled versions, the change-point is accepted. This makes the method non-parametric and robust to different types of noise.

For very long datasets, the authors introduce PCID$_W$, a windowed variant that splits the signal into manageable chunks to keep the computational cost of permutation testing low.

Datasets

The method was evaluated using:

  • Simulated Signals: 11 different scenarios (S1–S11) featuring various signal lengths, numbers of change-points, and noise types (von Mises, wrapped Cauchy, wrapped Normal, and serially correlated AR(1) noise).
  • Flare Data: 60 observations of the colatitude of illumination flares, a classic benchmark in circular statistics.
  • Acrophase Data: 306 observations of the timing of peak systolic blood pressure in a patient.
  • Wave Data: 1,326 observations of wave directions in the Adriatic Sea. This is the first time this specific dataset has been analyzed using change-point detection methods.

Key Insights and Ideas

  • The Power of Isolation: The "aha moment" here is that isolation makes detection easier. By ensuring a window contains at most one change-point, the contrast function's signal-to-noise ratio is maximized.
  • Robustness via Permutation: Even though the contrast function is mathematically derived assuming von Mises noise, the permutation test acts as a safety net. It allows the algorithm to perform well even when the noise is "heavy-tailed" (like the wrapped Cauchy distribution) or follows other patterns.
  • Handling Correlation: Real-world data is often "autocorrelated" (the current value depends on the previous one). The authors show that by subsampling the data and using a majority-voting rule across subsequences, PCID can accurately find changes even in "messy" time-series data.

Key Results

  • High Accuracy: In simulations, PCID achieved an Adjusted Rand Index (ARI)—a measure of similarity between true and estimated segments—frequently exceeding 0.99, indicating near-perfect detection.
  • Localization: The scaled Hausdorff distance (measuring how far estimated change-points are from true ones) was consistently near zero.
  • Real-World Utility:
    • On Flare data, PCID perfectly matched historical results (detecting changes at times 12 and 42).
  • On Wave data, PCID identified 68 change-points, providing a much more detailed "segmentation" of sea states than previous Hidden Markov Models, which only categorized the data into a few broad states.

Limitations and Future Work

  • Computational Cost: Permutation testing is inherently slower than purely analytical methods. While the windowed PCID$_W$ helps, very high-frequency data might still pose a challenge.
  • Piecewise Constant Assumption: The model assumes the signal jumps between constant values. It is not yet optimized for detecting changes in slopes or more complex functional forms on the circle.
  • Future Directions: The authors suggest extending this logic to toroidal data (data on a doughnut shape, like two angles recorded simultaneously) or cylindrical data (an angle plus a linear value, like wind direction and wind speed).

Summary updated

Abstract

In this paper we propose a new method for multiple change-point detection for piecewise-constant circular signals, a setting that, despite its importance in many scientific domains, remains comparatively under-explored. The proposed method, Permutation-based Circular Isolate-Detect, denoted PCID, uses an appropriately chosen contrast function and permutation testing to detect change-points in an offline manner, for the data sequence under consideration. Prior to detection, PCID isolates the change-points. The contrast function used is derived under the assumption of von Mises distribution for the noise, but we show that the method is robust and performs well for other distributions as well. Simulations are used to showcase the usability of the method in different signal and noise structures, including serially correlated noise. In order to exhibit the practical relevance of the method in real-world applications, PCID is applied to three real-world datasets, namely flare, acrophase and wave data.
BibTeX
@misc{loizidou2026multiple,
  author = {Sophia Loizidou and Andreas Anastasiou and Christophe Ley},
  title = {{Multiple change-point detection on the circle via isolation using permutation testing}},
  year = {2026},
  eprint = {2603.10772},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.10772},
}
Figure