Source-linked AI summary
Model-free inference of direct network interactions from nonlinear collective dynamics
Jose Casadiego, Mor Nitzan, Sarah Hallerberg, Marc Timme
TL;DR
Direct-interaction inference from nonlinear network dynamics typically requires a known model, limiting analysis when the governing dynamics are unknown. This paper introduces a model-free framework using explicit dependency matrices, basis-function expansions, and block-orthogonal regression, and reports reliable reconstruction across diverse dynamical regimes, interaction orders, and partial-observation settings.
Problem
The paper addresses how to identify direct pairwise and higher-order interactions from collective nonlinear dynamics without assuming a model in advance.
Method
The framework uses explicit dependency matrices, suitable basis-function expansions, and grouped-variable regression solved by the Algorithm for Revealing Network Interactions.
Results
The approach reconstructs direct interactions across transient, periodic, non-periodic, noisy, chaotic, and partially observed dynamics, including higher-order interactions.
Takeaways & Limitations
The framework supports model-free reconstruction from collective dynamics using short time series, noisy data, and systems with inaccessible units.
Takeaways & Limitations
Successful reconstruction requires sufficiently informative dynamics; synchronized or otherwise lower-dimensional states can make interaction existence and directionality impossible to extract.
Abstract
from arXiv · showhide
The topology of interactions in network dynamical systems fundamentally underlies their function. Accelerating technological progress creates massively available data about collective nonlinear dynamics in physical, biological, and technological systems. Detecting direct interaction patterns from those dynamics still constitutes a major open problem. In particular, current nonlinear dynamics approaches mostly require to know a priori a model of the (often high dimensional) system dynamics. Here we develop a model-independent framework for inferring direct interactions solely from recording the nonlinear collective dynamics generated. Introducing an explicit dependency matrix in combination with a block-orthogonal regression algorithm, the approach works reliably across many dynamical regimes, including transient dynamics toward steady states, periodic and non-periodic dynamics, and chaos. Together with its capabilities to reveal network (two point) as well as hypernetwork (e.g., three point) interactions, this framework may thus open up nonlinear dynamics options of inferring direct interaction patterns across systems where no model is known.
Mapping time series to direct interactions.
The framework infers direct pairwise and higher-order interactions from multivariate nonlinear dynamics without requiring a known dynamical model. It maps states to rates of change, decomposes interactions into basis functions, and identifies dependency blocks through structured regression.
- Problem formulation: The task is to identify which variables directly act on each unit, including pairwise links and higher-order interactions in which multiple units jointly influence one target.
- Dependency structure: Explicit dependency matrices encode which variables directly control each rate of change while treating pairwise and higher-order interactions uniformly.
- Dynamics-space representation: The model maps each system state to a unit’s rate of change and represents the resulting state-rate pairs as a smooth manifold in dynamics space.
- Structured regression: Unknown dynamics are decomposed first by interaction order and then into basis functions, producing a linear regression system whose coefficient blocks represent absent or existing interactions.
- Structured regression: The basis functions need only span a relevant function space; they need not exactly match the unknown interaction functions or yield a sparse representation.
- Structured regression: ARNI applies a greedy Block Orthogonal Least Squares procedure to multivariate time series and returns interactions ranked by their reduction of a fitting cost.
Revealing direct links in model systems.
The authors test the approach on transient, periodic, and non-periodic model dynamics using short trajectories combined across initial conditions. The experiments compare ARNI with correlation-based and information-theoretic interaction measures.
- Evaluation settings: The evaluation covers transient dynamics, periodic dynamics, and non-periodic oscillator dynamics, comparing ARNI with correlations, partial correlations, and transfer entropy.
- Sampling: Longer recordings and compositions of short trajectories improve predictions, indicating that sampling sufficient parts of state space is important for recovering direct interactions.
- Evaluation settings: For transients toward steady states, five-point trajectories from different initial conditions are combined into composite time series with M = S × m measurements.
- Evaluation settings: The transient benchmark includes a Michaelis–Menten network of N = 100 units with ni = 10 incoming connections per unit.
- Evaluation settings: The non-periodic benchmark uses phase-coupled oscillator networks, while the broader experiments also include biological systems with hypernetwork interactions.
Performance.
Performance remains reliable across network size, connectivity, noise, higher-order interactions, and partially observed systems. Reconstruction requires more data with greater noise or more hidden units, while longer time series improve quality.
- Scaling and robustness: The number Mθ of measurements needed to exceed an AUC threshold scales sublinearly with network size and linearly with the number of incoming connections per unit.
- Scaling and robustness: For N = 1000 and ni = 10, inferring one unit’s incoming connections takes 65 ± 26 s on conventional hardware.
- Scaling and robustness: Mθ increases supralinearly with noise level, but longer recordings improve reconstruction quality and preserve viability under highly noisy dynamics.
- Scaling and robustness: Reconstruction quality is independent of the probability of hypernetwork interactions because pairwise and higher-order interactions are decomposed and treated equally.
- Scaling and robustness: Links among measured units can remain reliably inferable with hidden units, although quality decreases as more units are hidden and improves with longer sampling.
Proper basis functions and learning curves.
Appropriate basis-function classes are sufficient to recover connectivity without exactly reproducing full dynamics. Learning curves identify the incoming-interaction count and the directly acting units when the class is appropriate.
- Choosing the correct interaction class is vital, although selecting the exact coupling function is unnecessary for connectivity reconstruction.
- The fitting cost is tracked as interactions are added, using recorded dynamics divided into training and validation sets.
- An L-shaped cost curve reveals the incoming-connection count at its knee, while the first interactions selected identify the directly acting units.
- With sufficiently short sampling intervals, time derivatives and their estimators can be obtained directly from recorded dynamics without model assumptions.
- Basis functions need only capture the interactions’ essential structure, not the full dynamics, to reveal network connectivity without a preknown model.
Effects of noise and hidden units.
Noise and hidden units reduce inference quality, but larger sampling collections retain useful information and generally outperform correlation-based alternatives. Figure 4 organizes performance across network size, connectivity, noise, hypernetwork structure, and observed-unit fraction.
- The experimental setting includes Gaussian-noise transients toward steady states with only a randomly selected subset of units recorded.
- Figure 4 evaluates required time-series length against network size, incoming interactions, and noise, and examines hypernetwork interactions and hidden units.
- Noise and hidden units moderately reduce performance, while inference quality still increases with the number of observations.
- With incomplete observations, the approach generally outperforms correlations, partial correlations, and transfer entropy across different recorded-unit fractions.
Robust inference of biological networks.
The framework reconstructs interactions in model biological systems from transient dynamics toward periodic orbits. In these demonstrations, more observations improve predictions and reconstruction quality exceeds correlation-based methods.
- Table 1 presents alternative basis-function representations for interactions.
- The framework reconstructs interactions in yeast glycolytic oscillators and Drosophila circadian-clock models from transient dynamics toward periodic orbits.
- It combines a dynamics-space representation, suitable basis-function expansions, and orthogonal-least-squares regression.
- Larger numbers of observations improve reconstruction predictions in both biological model systems.
- Reconstruction quality again outperforms correlations, partial correlations, and transfer entropy.
Discussion
The framework reconstructs direct interactions from nonlinear dynamical data without prior models, while remaining robust across varied regimes and scalable to larger networks. Its sampling requirements are minimal, but low-dimensional or synchronized dynamics can limit recoverability.
- Framework: The framework infers direct network and hypernetwork interactions from nonlinear collective dynamics without requiring prior coupling functions or sparse representations.It uses dependency matrices and grouped-variable regression to separate pairwise, three-point, and higher-order influences.
- Robustness: Reconstructions remain robust across diverse dynamical regimes, pairwise and hypernetwork interactions, noise, and hidden units.The framework is also described as applicable to essentially arbitrary nonlinear dynamics and partially inaccessible systems.
- Sampling: Short, potentially heterogeneous time series can suffice, and controlled external driving is generally unnecessary for transients or periodic dynamics.Transient, stochastic, or otherwise sufficiently complex dynamics can provide more interaction information than stable low-dimensional trajectories.
- Scalability: The number of independent measurements grows linearly with the local number of interaction partners and sublinearly with network size.This scaling supports reconstruction of large systems and can be paired with learning curves when ground truth is unavailable.
- Scope boundary: Synchronized states and other low-dimensional invariant dynamics may not provide full information about network interactions.In synchronized states, interaction existence and directionality can be impossible to extract from time series.
Methods
The methods evaluate reconstruction across biological, oscillatory, non-periodic, and chaotic network models, including higher-order interactions and noisy signals. Biological examples include yeast glycolytic oscillations and the Drosophila circadian clock.
- Model systems: The evaluation includes transient dynamics toward steady states generated with Michaelis–Menten kinetics, commonly used for gene-regulation models.These systems contain randomly selected incoming connections per node, with weighted directed links represented by J_ij.
- Biological models: Biological periodic systems include the yeast glycolytic oscillator and the Drosophila circadian clock, both containing hypernetwork interactions.In hypernetworks, two units jointly and directly influence a third.
- Dynamical regimes: Non-periodic dynamics are generated with phase-coupled oscillators whose coupling functions contain two Fourier modes.The model uses constant natural frequencies and coupling terms involving first- and second-harmonic sine functions.
- Higher-order interactions: The hypernetwork extension introduces a second-order interaction matrix E_i whose elements quantify joint direct influence of units j and k on unit i.This explicitly represents three-point interactions in addition to pairwise interactions.
- Dynamical regimes: Chaotic dynamics are generated with networks of coupled Rössler oscillators, whose oscillator states are three-dimensional and can receive external noisy signals.Each oscillator is defined by three differential equations.