🌀 nonlinear sciences

Quantifying the complexity of trajectory ensembles with clustering-weighted multivariate multiscale sample entropy

This paper introduces clustering-weighted multivariate multiscale sample entropy (CWMMSE), a novel metric that quantifies the complexity of trajectory ensembles by distinguishing between individual dynamical complexity and population diversity, thereby overcoming the limitations of traditional averaging methods across diverse physical, environmental, and biomedical systems.

Original authors: Chenxiao Tian, J/"urgen Hackl

Published 2026-07-17

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Chenxiao Tian, J/"urgen Hackl

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ 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

The Symphony of Chaos: Why One Voice Isn't the Whole Story

Imagine you are trying to understand the weather. You could look at a single thermometer in your backyard and say, "It's hot today." But that doesn't tell you if a storm is brewing, if the wind is howling, or if the whole region is in a state of chaotic turbulence. In the world of science, from tracking satellites in space to monitoring heartbeats in a hospital, researchers often deal with "ensembles." This is just a fancy word for a large group of moving things, like a fleet of ships, a swarm of birds, or a crowd of patients. For a long time, scientists tried to understand these groups by measuring just one thing: the average. They would take the "complexity" (how unpredictable or messy the movement is) of every single member, add them up, and divide by the number of members. It's like judging a choir by the average volume of all the singers.

But here is the problem with averages: they hide the truth. Imagine two different choirs. In the first, every single singer is singing a wildly complex, difficult jazz solo, but they are all singing the exact same solo. In the second, every singer is humming a simple, boring tune, but they are all humming different tunes. If you just take the average, both choirs might look the same. But they are totally different! The first choir is redundant (boring because everyone is the same), and the second is diverse (interesting because everyone is different). To truly understand a system, you need to know not just how messy the individual parts are, but how varied the whole group is. This paper tackles the question of how to measure the complexity of a whole crowd, not just the average person in it.

The New Tool: A "Weighted" Score for Crowds

The authors of this paper, Chenxiao Tian and Jürgen Hackl from Princeton University, have invented a new mathematical tool called Clustering-Weighted Multivariate Multiscale Sample Entropy (CWMMSE). That is a mouthful, so let's break it down with a simple analogy.

Think of a busy city street. You have thousands of people walking around.

The First Ingredient (Individual Complexity): First, you look at one person. Are they walking in a straight line, or are they zig-zagging wildly, dodging obstacles, and changing direction constantly? If they are zig-zagging, they have high "individual complexity."
The Second Ingredient (Ensemble Diversity): Next, you look at the whole crowd. Are they all walking in the exact same zig-zag pattern (like a marching band)? Or is everyone doing something totally different? If everyone is doing the same thing, the crowd is "redundant." If everyone is doing something unique, the crowd is "diverse."

The old way of measuring the city was to just average the zig-zagging of everyone. The authors say this is wrong. Their new tool, CWMMSE, does two things: it groups people who are walking similarly into "clusters" (patterns), and then it calculates a score that rewards a crowd only if it is both individually messy and collectively diverse.

The Big Discovery: The Average Can Lie

The paper proves that the old "average" method often gets the answer completely backwards. The authors tested their new tool on eleven different real-world systems, from ocean currents to heartbeats, and found some surprising results.

The Ocean Test:
They looked at two groups of ocean drifters (floating buoys). One group was in the Gulf Stream, a super-fast, energetic river of water. The other was in the Sargasso Sea, a calm, quiet area.

The Old Way: The drifters in the Gulf Stream were individually more chaotic (they were getting tossed around by eddies). So, the average said the Gulf Stream was the "more complex" system.
The New Way (CWMMSE): The authors found that in the Gulf Stream, all the drifters were actually following the same few giant loops and swirls. They were individually messy, but collectively, they were all doing the same thing. The Sargasso Sea drifters, however, were moving slowly but each one was taking a totally unique, random path.
The Result: CWMMSE declared the calm Sargasso Sea the more complex system because it was more diverse. The energetic Gulf Stream was actually "redundant."

The Heartbeat Test:
This is where it gets really important for medicine. They looked at heartbeats from healthy people and people with heart disease (myocardial infarction).

The Old Way: When they averaged the complexity of individual heartbeats, the sick hearts actually looked more complex (more irregular) than the healthy ones. This is a known problem in medicine; sometimes disease makes a signal look "noisier," but that doesn't mean the system is healthy. In fact, the average method ranked the diseased hearts as the more complex group, which was the wrong conclusion.
The New Way (CWMMSE): They found that healthy hearts, while individually slightly less complex than the sick ones, showed a huge variety of different patterns across the whole group of people. Sick hearts, however, all fell into a few stereotyped, "broken" patterns. The healthy group was diverse; the sick group was redundant.
The Result: CWMMSE correctly identified the healthy group as the more complex, robust system. It reversed the ranking from the average method, showing that the "messiness" of the sick hearts was actually a sign of a collapsing, less diverse population.

What This Rules Out

The authors are very clear about what their tool is not. They explicitly show that you cannot just measure "diversity" alone. They tested this with a group of satellites and space debris. These objects are all in different orbits, so they are incredibly diverse (high diversity score). But their movement is perfectly predictable and simple (low individual complexity).

The Trap: If you only measure diversity, you would say space debris is the most complex system in the universe.
The Reality: CWMMSE correctly gives this group a score near zero. It proves that a system can be diverse but still be "simple" if the individual parts are predictable. You need both ingredients to get the right answer.

How Sure Are They?

The authors didn't just guess; they tested this rigorously. They used computer simulations of chaotic systems (like the famous Lorenz and Rössler systems) where they knew the answer in advance. Their new tool got the ranking right every time, while the old average method failed. They also ran "null tests" (like the satellite example) to make sure their tool didn't get tricked by things that looked complex but weren't.

They tested this on eleven different types of data, including ocean currents, tropical cyclones, ship traffic, earthquakes, and even the gait of people with neurodegenerative diseases. In every single case, the new tool revealed a hidden structure that the old average method missed. They even showed that the "coupling" between individual messiness and group diversity matters; it's not just a simple multiplication of the two, but a specific relationship that their formula captures.

The Takeaway

The main lesson here is that the whole is not just the sum of its parts. When scientists look at a crowd of data points, they shouldn't just average them out. Sometimes, a calm, diverse group is more complex and interesting than a chaotic, repetitive one. By using this new "weighted" score, researchers can finally tell the difference between a system that is truly complex and one that is just messy or just different. It's a new way to listen to the whole choir, rather than just the average volume of the singers.

Technical Summary: Quantifying the Complexity of Trajectory Ensembles with Clustering-Weighted Multivariate Multiscale Sample Entropy (CWMMSE)

1. Problem Statement

In physical, biological, and engineering sciences, systems are increasingly observed as ensembles of trajectories (e.g., satellite constellations, clinical cohorts, fluid particles) rather than single time series. Existing complexity measures, such as Sample Entropy (SampEn), Approximate Entropy, and Multiscale Entropy (MSE), characterize the unpredictability of individual time series. However, applying these measures to ensembles by simply averaging the complexity of individual trajectories discards the population's structural diversity.

The paper identifies a critical limitation in this "averaging" approach: it conflates individual complexity (how irregular a single trajectory is) with ensemble diversity (how varied the population of trajectories is). Consequently, averaging cannot distinguish between:

A redundant population where every member exhibits the same intricate behavior.
A diverse population where many distinct, individually simple behaviors coexist.

The authors argue that a faithful measure of system complexity must simultaneously account for the irregularity of members and the diversity of the population, as these two components can diverge significantly in real-world data.

2. Methodology: Clustering-Weighted Multivariate Multiscale Sample Entropy (CWMMSE)

The authors propose CWMMSE, a metric that fuses individual complexity and ensemble diversity into a single weighted entropy value. The construction proceeds in three stages:

A. Within-Trajectory Complexity (Individual Component)
For each trajectory $T_j$ in an ensemble of $M$ multivariate time series, the authors compute the Multivariate Multiscale Sample Entropy (MMSE).

Coarse-graining: At temporal scale $s$ , channels are averaged over non-overlapping blocks.
Delay Embedding: Composite delay vectors are formed across all channels.
Entropy Calculation: The conditional probability that two segments matching at embedding dimension $d$ also match when extended by one sample per channel is calculated. The negative log of this probability yields $S(T_j)$ .
This step captures the intrinsic dynamical complexity of each individual trajectory.

B. Ensemble Diversity (Population Component)
The trajectories are grouped into behavioral patterns via clustering.

Clustering: Agglomerative Ward linkage is used on a trajectory dissimilarity metric (default: point-wise Euclidean distance).
Resolution: The clustering is cut at a specific height (defined as a fraction of the maximum linkage distance) to define behavioral patterns.
Diversity Metric: The population distribution over these $k$ clusters is quantified using the Shannon cluster entropy, $H(C) = -\sum p_i \ln p_i$ , where $p_i$ is the fraction of trajectories in cluster $i$ . This is normalized by $\ln M$ to yield $\hat{H}$ .

C. Fusion: The Weighted Entropy
CWMMSE combines these ingredients using the weighted entropy framework of Belis and Guiaşu.

Formula: $S_{CWMMSE} = \sum_{i=1}^{k} (-p_i \ln p_i) S_i$ , where $S_i$ is the mean MMSE of trajectories in cluster $i$ .
Interpretation: This is the Shannon entropy of the population's pattern distribution, where each outcome (pattern) is weighted by its own dynamical complexity.
Theoretical Basis: The authors prove that the empirical plug-in estimator of CWMMSE is strongly consistent for a fixed finite partition as the ensemble size $M \to \infty$ .

Key Distinction from Alternatives:

Vs. Averaging: Averaging ( $\bar{S}$ ) ignores diversity.
Vs. Diversity Indices (e.g., Rao's Quadratic Entropy): These weight by pairwise dissimilarity, failing to distinguish between a diverse population of complex behaviors and a diverse population of trivial (predictable) behaviors.
Vs. Separable Product ( $H(C) \cdot \bar{S}$ ): While the product respects the limits of both ingredients, it discards the complexity-diversity coupling. CWMMSE retains the covariance between a pattern's complexity and its prevalence (surprisal), which is non-negligible in real data.

3. Key Contributions

Definition and Theory: Introduction of CWMMSE as a specific instance of Belis and Guiaşu weighted entropy, applied to trajectory ensembles, with a proof of strong consistency for the empirical estimator under fixed resolution.
Decomposition: Demonstration that system complexity decomposes into two divergent components: within-trajectory complexity and between-trajectory diversity.
Necessity of Dual Ingredients: Validation through two "controls" showing that single-ingredient measures fail:
- Cardiac Control: Averaged entropy incorrectly ranks diseased hearts as more complex than healthy ones; CWMMSE correctly identifies the diseased cohort as less complex due to a collapse in population diversity.
- Orbital Null: Diversity-only indices (like Rao's entropy) rank a cloud of perfectly predictable satellite debris as highly complex; CWMMSE correctly returns near-zero complexity because the individual trajectories are trivial.
Cross-Domain Validation: Application of the measure across eleven systems (chaos, ocean drifters, ECGs, satellite orbits, turbulence, cyclones, etc.) without per-system tuning, showing the decomposition holds universally.

4. Results

Deterministic Chaos: CWMMSE correctly orders regimes of Lorenz and Rössler systems, peaking in chaotic regimes and approaching zero in fixed-point/periodic regimes. It distinguishes nonlinear chaos from linear noise surrogates.
Ocean Drifters: The Gulf Stream (energetic flow) has higher individual complexity (tortuous paths) but lower system complexity (redundant, coherent structures) compared to the Sargasso Sea (quiescent gyre), which has lower individual complexity but higher system complexity due to diverse, independent paths. CWMMSE correctly ranks the Sargasso as the more complex system.
ECG Cohorts: In myocardial infarction (MI) vs. healthy cohorts, individual per-record complexity is nearly identical (or slightly higher in MI). However, the healthy cohort exhibits significantly higher ensemble diversity (more varied morphologies). CWMMSE reverses the ranking of averaged entropy, correctly identifying the healthy population as more complex.
Satellite Orbits: A debris cloud of low-Earth-orbit objects is maximally diverse in position but individually predictable. Diversity-only indices rank this as the most complex system; CWMMSE correctly assigns it near-zero complexity.
Other Systems: The measure successfully distinguishes turbulent from laminar flow, orders tropical cyclone tracks by erraticism, and detects a collapse in system complexity during the 2011 Tohoku earthquake wave train.

5. Significance and Claims

The paper claims that population complexity should be measured rather than averaged. The central argument is that the complexity of an ensemble is not merely the sum or average of its parts; it is a distinct property arising from the interaction between individual irregularity and population diversity.

Theoretical Significance: CWMMSE provides a principled, information-theoretic framework (weighted entropy) that unifies the concepts of dynamical complexity and functional diversity. It resolves the ambiguity where a system can be "complex" in one sense (individual trajectories) but "simple" in another (redundant population), or vice versa.
Practical Significance: The measure offers a diagnostic tool for fields ranging from physiology (detecting disease via loss of population diversity) to space situational awareness (distinguishing complex debris from predictable orbits).
Limitations and Scope: The authors modestly note that CWMMSE is a comparative instrument within a study, dependent on clustering resolution and ensemble size. It ranks systems of comparable stochasticity rather than separating structure from noise (as sample entropy rates noise as highly irregular). The resolution parameter is treated as an explicit, reportable setting rather than a nuisance to be eliminated.

In conclusion, the paper establishes that for systems observed as trajectory ensembles, a faithful measure of complexity must capture both the intricacy of the members and the variety of the population, a requirement met by the clustering-weighted multivariate multiscale sample entropy.