Distinguishing pairwise and higher-order interactions in coupled oscillators from time series

This paper introduces an adaptive LASSO-based method that accurately distinguishes between pairwise and higher-order interactions in coupled oscillator systems from time series data, outperforming existing baselines in synthetic tests and demonstrating practical utility in analyzing human brain network dynamics.

The Gist

Imagine you are standing in a crowded room where everyone is clapping. Sometimes, people clap in perfect unison. Sometimes, they clap in small groups. Sometimes, the rhythm is chaotic.

Your goal is to figure out how they are coordinating. Are they just listening to the person right next to them? Or are they listening to a specific trio of people? Or is it a mix of both?

This is exactly the problem scientists face when studying coupled oscillators. These are systems found everywhere in nature: heartbeats, neurons firing in the brain, fireflies flashing, or even the rhythmic swaying of a bridge. They all "oscillate" (move back and forth in a cycle).

The paper by Weiwei Su and colleagues tackles a tricky question: Can we look at the "clapping" (the time-series data) and figure out if the coordination is happening between pairs of people, groups of three, or a mix of both?

Here is a simple breakdown of their solution and why it matters.

The Problem: The "Look-Alike" Illusion

The authors point out a major headache: Different causes can look like the same result.

Imagine two scenarios:

1. Scenario A: Everyone is clapping because they are listening to their immediate neighbor (Pairwise interaction).
2. Scenario B: Everyone is clapping because they are listening to a specific group of three friends (Three-body interaction).

If you just watch the crowd for a while, the overall rhythm (the "order parameter") might look almost identical in both cases. It's like trying to guess the ingredients of a soup just by tasting the final broth; you might taste "salt," but you can't tell if it came from salt shaker A or salt shaker B.

In the past, scientists had to guess. They often assumed everything was just pairs (A talking to B), which led to wrong conclusions about how complex systems like the brain work.

The Solution: The "Adaptive LASSO" Detective

The team developed a new mathematical tool called Adaptive LASSO. To understand how it works, let's use an analogy.

Imagine you are a detective trying to solve a crime. You have a list of 100 suspects (the oscillators). You know that only a few of them actually committed the crime (the connections), but you don't know which ones.

• Old Method (OLS): This is like a detective who accuses everyone of being slightly suspicious. They say, "Well, Suspect #42 might have done it, and Suspect #99 might have helped a little bit." They can't distinguish between "no crime" and "a tiny crime." This leads to a lot of false alarms.
• Standard LASSO: This detective is better. They use a "shrinkage" technique to say, "Most of these suspects are innocent." But they are a bit too aggressive and sometimes accidentally clear the guilty or still miss the subtle clues.
• Adaptive LASSO (The New Method): This is the super-detective. They don't just shrink the list; they adapt their strategy based on the evidence.
  1. First, they make a quick guess of who is involved.
  2. Then, they apply a "smart filter." If the evidence for a connection is weak, they confidently say, "No, that's zero. They aren't connected." If the evidence is strong, they keep it.
  3. Crucially, this method is honest. It rarely accuses an innocent person (false positive) and rarely misses a guilty one (false negative).

What They Discovered

The team tested their "super-detective" on computer simulations and real-world data.

1. It Works on Fake Data: They created digital worlds where they knew exactly who was connected to whom. The new method correctly identified the "pair" vs. "trio" connections almost 100% of the time, while the old methods got confused and made many mistakes.
2. It Works on Real Brains: They applied the method to data from the human brain (90 different regions). They successfully mapped out how different parts of the brain talk to each other, distinguishing between simple two-way conversations and complex three-way group chats.
3. It Handles Noise: Real life is messy. People cough, the wind blows, and data is noisy. The old methods crumbled under noise, but the Adaptive LASSO kept its cool and found the true connections even when the data was "noisy."

Why This Matters

Understanding the difference between pairwise (two-way) and higher-order (group) interactions is like understanding the difference between a simple phone call and a complex group meeting.

• In Medicine: If we can tell if brain regions are talking in pairs or groups, we might spot early signs of diseases like epilepsy or Alzheimer's, where these communication patterns break down.
• In Nature: It helps us understand how fireflies synchronize or how heart cells beat together without a conductor.

The Bottom Line

This paper gives us a powerful new lens to look at the world. Instead of just seeing a crowd clapping, we can now figure out the hidden rules of who is listening to whom. The authors have built a tool that cuts through the noise and the confusion to reveal the true structure of complex, rhythmic systems.

In short: They found a way to stop guessing whether a system is a collection of pairs or a collection of groups, and they did it with a mathematical tool that is sharper and more reliable than anything we had before.

Technical Summary

1. Problem Statement

Rhythmic phenomena in biological systems (e.g., brain activity, heartbeats) are often modeled as networks of coupled limit-cycle oscillators. A critical challenge in analyzing these systems is distinguishing between pairwise (two-body) interactions and higher-order (e.g., three-body) interactions.

• The Core Difficulty: Different interaction types can produce nearly identical macroscopic dynamical states (e.g., synchronization levels, cluster states) and order parameters, making it impossible to distinguish them solely by observing the system's collective behavior.
• The Gap: Existing methods for network reconstruction from time series data (such as Ordinary Least Squares or standard LASSO) generally assume pairwise interactions or fail to accurately differentiate between zero coupling and weak coupling in noisy environments.
• Objective: The authors aim to develop a method to infer not only the topology (who interacts with whom) and strength of couplings but also to identify the type of interaction (pairwise vs. three-body vs. mixed) purely from phase time-series data, even in the presence of noise.

2. Methodology

The authors propose a method based on Adaptive LASSO (Least Absolute Shrinkage and Selection Operator) to infer interaction networks from phase time series $\{\phi_i(t)\}$.

A. Mathematical Model

The system is modeled as a population of $N$ oscillators governed by the following stochastic differential equation:
$$\frac{d\phi_i}{dt} = \omega_i + f_i^{(2)}(\vec{\phi}; W_i^{(2)}) + f_i^{(3)}(\vec{\phi}; W_i^{(3)}) + \sigma_i \xi_i(t)$$

• $\omega_i$: Natural frequency.
• $f_i^{(2)}$: Pairwise coupling function (sinusoidal).
• $f_i^{(3)}$: Three-body coupling function (symmetric simplicial interaction).
• $W$: Coupling strength vectors (parameters to be estimated).
• $\xi_i(t)$: Gaussian white noise.

B. The Adaptive LASSO Algorithm

The estimation process involves two main steps for each oscillator $i$:

1. Preliminary Estimation (OLS): The parameters (natural frequency and coupling strengths) are first estimated using Ordinary Least Squares (OLS) to minimize the squared error between the observed phase derivative and the model prediction.
   $$E_{OLS} = \sum \left( \frac{\Delta \phi_i}{\Delta t} - \omega_i - f^{(2)} - f^{(3)} \right)^2$$
2. Refinement (Adaptive LASSO): The preliminary estimates are used to define adaptive weights ($c_{ij} = 1/\tilde{W}_{ij}$). A cost function is minimized that includes an $L_1$ penalty weighted by these adaptive factors:
   $$E_{AL} = E_{OLS} + \alpha \left( \sum c_{ij}^{(2)} |W_{ij}^{(2)}| + \sum c_{ijl}^{(3)} |W_{ijl}^{(3)}| \right)$$
   • Optimization: Solved using Least Angle Regression (LARS).
   • Hyperparameter Selection: The regularization parameter $\alpha$ is selected by minimizing the Bayesian Information Criterion (BIC).
   • Mechanism: This approach satisfies "oracle properties," meaning it can consistently select the correct non-zero parameters (identifying true couplings) and estimate their values accurately, effectively shrinking noise-induced weak couplings to zero.

C. Interaction Type Classification

To determine if the system is dominated by pairwise, three-body, or mixed interactions, the authors employ a Z-test:

• Calculate the proportion of non-zero inferred couplings for pairwise ($\hat{q}^{(2)}$) and three-body ($\hat{q}^{(3)}$) terms.
• Test the null hypothesis $H_0: q^{(2)} = q^{(3)}$.
• If $H_0$ is rejected, the interaction type with the higher probability is deemed dominant. If not rejected, the system is classified as a mixture.

3. Key Contributions

1. Novel Algorithm: Introduction of Adaptive LASSO specifically tailored for coupled oscillator systems to distinguish between pairwise and higher-order interactions, a task where standard LASSO and OLS fail.
2. Robustness to Noise: The method successfully distinguishes interaction types and infers topologies in noisy environments where macroscopic order parameters (like the Kuramoto order parameter $r$) are indistinguishable between different interaction types.
3. Extension to General Systems: The framework is extended beyond simple Kuramoto models to include:
   • Oscillators with phase-dependent amplitudes (limit cycle deformations).
   • Systems with four-body interactions.
4. Real-World Application: Successful application to human brain network data (structural connectivity from 90 cortical regions), demonstrating practical utility in complex biological systems.

4. Results

The method was evaluated using synthetic datasets and compared against two baselines: OLS (with statistical testing) and standard LASSO.

• Topology Inference Accuracy:
  • Adaptive LASSO achieved 100% accuracy in identifying the correct interaction type (pairwise, three-body, or mixed) and inferring the network topology (True Positives) with zero False Positives and zero False Negatives in synthetic tests.
  • Standard LASSO produced false positives (incorrectly inferring three-body couplings in pairwise-only systems) and underestimated coupling strengths.
  • OLS produced a high number of false positives and overestimated coupling strengths.
• Quantitative Metrics:
  • Matthews Correlation Coefficient (MCC): The proposed method consistently achieved higher MCC values (approaching 1.0) compared to baselines across various coupling probabilities ($q=0.05, 0.10, 0.15$) and observation durations.
  • Mean Squared Error (MSE): The proposed method yielded the lowest MSE for coupling strength estimation, whereas OLS overestimated and LASSO underestimated strengths.
• Observation Time: Accurate inference (MCC > 0.8) required longer observation periods for denser networks (e.g., >100 cycles for $q=0.15$) compared to sparse networks.
• Real-World Application (Human Brain): Applied to 90-node brain networks. The method accurately reconstructed strong couplings with an overall MCC of 0.694 and a very low MSE ($1.15 \times 10^{-6}$) for strength estimation, outperforming LASSO and OLS.
• Higher-Order Extensions: The method successfully inferred four-body interactions (MCC 0.782), though performance decreased slightly compared to three-body cases, suggesting a need for longer time series for very high-order terms.

5. Significance

• Theoretical Advancement: This work resolves a fundamental ambiguity in network science: the inability to distinguish interaction orders based solely on dynamical states. It provides a rigorous statistical framework for "unmixing" interaction types from time-series data.
• Biological Relevance: By successfully applying the method to human brain data, it opens new avenues for understanding the role of higher-order interactions (simplicial complexes) in neural synchronization, potentially leading to new biomarkers for neurological diseases.
• Methodological Impact: The demonstration that Adaptive LASSO outperforms standard regression and LASSO in this specific context suggests that adaptive weighting is crucial for sparse network reconstruction in noisy, non-linear dynamical systems.
• Future Directions: The authors highlight that while the method works for 4-body interactions, computational costs scale exponentially ($O(N^k)$). Future work will focus on developing more efficient algorithms for large-scale, high-order network inference.

In conclusion, the paper presents a robust, mathematically grounded tool that transforms the analysis of coupled oscillators from merely observing synchronization patterns to inferring the underlying mechanistic structure of interactions, distinguishing between simple pairwise links and complex group dynamics.