Efficient parameter inference in networked dynamical systems via steady states: A surrogate objective function approach integrating mean-field and nonlinear least squares

Yanna Ding, Malik Magdon-Ismail, Jianxi Gao · 2025

Read Efficient parameter inference in networked dynamical systems via steady states: A surrogate objective function approach integrating mean-field and nonlinear least squares by Yanna Ding, Malik Magdon-Ismail, Jianxi Gao (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-18335.txt.

Introduction

Yanna Ding, Malik Magdon-Ismail, and Jianxi Gao present a 2025 paper on inferring the rules of a networked system from a single snapshot of its resting state. Gene levels, species counts, and infection tallies often settle into steady states, and only noisy versions of those are measurable. Estimating the governing parameters the usual way means simulating coupled differential equations again and again, which is slow.

Their surrogate objective swaps the coupled system for decoupled equations that yield steady states cheaply. Fitting that surrogate also denoises the observations, landing closer to the truth than the raw data. The fitted model even predicts new equilibria if the network's wiring changes. Tests on ecological, gene regulatory, and epidemic networks confirm the gains. The authors offer the method for the many fields where time series simply do not exist.

Catalog record

CREATORYanna Ding, Malik Magdon-Ismail, Jianxi Gao
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE52 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-18335

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Efficient parameter inference in networked dynamical systems via steady states: A surrogate objective function approach integrating mean-field and nonlinear least squares Yanna Ding, Malik Magdon-Ismail, Jianxi Gao Abstract In networked dynamical systems, inferring governing parameters is crucial for predicting nodal dynamics, such as gene expression levels, species abundance, or population density. While many parameter estimation techniques rely on time-series data, particularly systems that converge over extreme time ranges, only noisy steady-state data is available, requiring a new approach to infer dynamical parameters from noisy observations of steady states. However, the traditional optimization process is computationally demanding, requiring repeated simulation of coupled ordinary differential equations (ODEs). To overcome these limitations, we introduce a surrogate objective function that leverages decoupled equations to compute steady states, significantly reducing computational complexity. Furthermore, by optimizing the surrogate objective function, we obtain steady states that more accurately approximate the ground truth than noisy observations and predict future equilibria when topology changes. We empirically demonstrate the effectiveness of the proposed method across ecological, gene regulatory, and epidemic networks. Our approach provides an efficient and effective way to estimate parameters from steady-state data and has the potential to improve predictions in networked dynamical systems. Preprint: APS/123-QED I Introduction Knowing a network’s dynamics is essential for understanding and predicting its behavior in complex systems, with broad application in ecology [ 1 , 2 , 3 , 4 , 5 , 6 ] , biology [ 7 , 8 , 9 , 10 ] , epidemiology [ 11 , 12 ] , energy supply [ 13 , 14 ] , neurophysiology [ 15 ] , and organizational psychology [ 16 ] . Network dynamics often consists of three fundamental components: dynamical models, parameters, and network topology. The dynamical models are closed-form mathematical expressions, including polynomials, trigonometric functions, etc. Dynamical parameters are essential numerical values, such as polynomial orders and scaling factors within these mathematical expressions. The network topology determines the coupling strengths between nodes. Our focus is parameter estimation, given underlying topology, dynamical models, and observed steady states. Accurate parameter estimation is critical for computing important network tasks, including computing tipping point leading to systems collapses [ 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 ] , network control [ 25 , 26 ] , synchronization [ 27 ] , and many more. There are existing methods for inferring parameters [ 28 , 29 ] , topology [ 30 ] , and dynamical models [ 31 , 32 , 33 , 34 ] . These approaches typically use a time series of system states, which is much richer information than just steady states. For inferring dynamical models, the two approaches are to use semi-parametric approaches like [ 31 , 32 , 33 , 34 ] or some deep learning methods [ 35 , 36 , 37 ] . Notably, these deep learning methods often require extensive time-series data for effective learning and may struggle to generalize to unobserved dynamical systems. Also note that existing methods relying on a time series of states can be costly or even impossible in certain scenarios [ 38 ] . Ecological systems evolve over millennia [ 39 , 40 ] , and we only observe the current steady state of each species. For biomedical networks, equilibration is at small time scales, making it difficult for experimental devices to capture interim states [ 10 , 41 ] . Despite recent progress to learn microscopic dynamics from multiple sets of steady states generated by perturbation experiments [ 38 , 42 , 43 ] , there is still a lack of a universal framework to infer dynamical parameters from a single noisy steady-state observation of a networked dynamics. Parameter estimation for differential equations is often time-consuming due to repeated numerical integration. One approach to address this issue is fitting proposed models to the derivatives of the observed time-series data approximated using finite difference [ 30 ] . However, this approach may yield inaccurate results when the data resolution is low, and, of course, it is not applicable when only the equilibrium state is available. One common objective function to estimate parameters is to minimize the discrepancy between observed states and those generated by the learned model using numerical integration techniques [ 29 ] . However, repeated numerical integration to get network states is computationally intensive, especially for coupled complex systems with nonlinear dynamics. In this paper, our aim is to improve the efficiency of parameter estimation for ODEs when the only available data is noisy observation of steady states. The mean-field network reduction approach, introduced in [ 44 ] , simplifies coupled systems while attempting to preserve their original dynamics. This approach collapses a high-dimensional network into a one-dimensional effective state which drives the dynamics of all other states. Extensive studies have been conducted to evaluate the accuracy of this approach [ 44 , 45 ] . Alternative dimensionality reduction methods have been explored, including those proposed in [ 46 , 47 , 48 ] . Furthermore, researchers have extended network reduction approaches to forecast network collapse [ 18 ] , predict steady states from incomplete network topology [ 49 ] and infer degrees from observed steady states [ 50 ] . While these previous works assume prior knowledge of dynamical parameters, our work builds upon the mean-field approach and combines it with nonlinear least squares (NLS) to estimate unknown parameters. In this work, we propose a surrogate objective function based on mean-field estimation of steady states that significantly reduces computational time while preserving the shape of the exact NLS objective function. The parameter-specific steady states are approximated using a low-dimensional, decoupled system of ODEs in the mean-field approximation. We demonstrate empirically that NLS inference using the surrogate objective produces parameters that recover steady states that match the ground truth more accurately than the observed noisy steady states. Furthermore, we demonstrate that the parameters inferred through our method can predict state evolution when the network undergoes changes. Figure 1 : Workflow Illustration on Epidemic Dynamical Systems. (a) Ground Truth Epidemic System and Observed Equilibrium: (a1) Ground truth steady states ( \mathbf{x}^{*} ) simulated using coupled ODEs with the ground truth parameter B^{*}=0.5 . (a2) Network topology induced by the adjacency matrix A , generated using the Erdős Rényi model ( N=100 , average degree 12 ). (a3) Comparison of observed states ( \mathbf{y} ) with true steady states. Observations are sampled from a normal distribution with mean x_{i}^{*} and standard deviation 10\%\lvert x_{i}^{*}\rvert . (b) Optimization of Surrogate Objective Function (Eq. ( 18 )): The optimization process takes as input mathematical expressions representing the self-dynamics and interaction terms, the degree of each node, and the observed steady states \mathbf{y} . A gradient-based parameter search algorithm is employed for optimization. Note that numerical integration is performed for each query parameter B^{\tau} during the optimization. The detailed computation of the gradients is provided in Appendix C . (c) Inferred Parameter \hat{B} and Predicted Equilibrium: (c1) State evolution of the ODE with the inferred parameter \hat{B} , resulting in the predicted equilibrium \hat{\mathbf{x}}=\mathbf{x}(\hat{B}) . (c2) Close alignment between predicted steady states and ground truth. II Parameter Estimation via a Surrogate Objective Function II.1 Problem

Read the complete text