Introduction
Identifying the underlying mechanisms of complex systems is a cornerstone of network science. While static network identification is well-studied, many real-world systems—ranging from power grids to biological pathways—exhibit time-varying structures. These Switching Network Dynamical Systems (SNDS) change their internal node dynamics and interaction topologies abruptly, often triggered by internal state shifts or external events. The fundamental challenge, known as system identification, involves reconstructing both the continuous dynamics and the discrete switching logic solely from observed node trajectories.
Problem Formulation
The identification of SNDS presents a "chicken-and-egg" dilemma: to determine the network topology, one must know the system's current mode; conversely, to identify the mode, the dynamics and topology must be known. This typically results in a non-convex, mixed-integer optimization problem that is computationally prohibitive for large systems. Traditional methods often require prior knowledge of switching times or labels, which are rarely available in practice.
Methodology: A Convex Optimization Framework
To address these challenges, we propose a framework that transforms the combinatorial search into a tractable sequence of convex programs. The approach relies on three key technical pillars:
- Polynomial Approximation: Unknown node dynamics are modeled using a polynomial basis, converting a functional search into a linear coefficient estimation problem.
- Convex Relaxation: Discrete switching constraints are relaxed using Moment-based Semidefinite Relaxation (SDR). This allows the optimizer to handle "soft" mode assignments that converge to discrete values.
- Alternating Optimization: The problem is solved by alternating between two steps:
- Mode Search: Fixing the dynamics and graphs to solve a Semidefinite Program (SDP) for the switching signal.
- Dynamics/Network Search: Fixing the switching signal to solve a Linear Program (LP) that refines node equations and adjacency matrices.
Experimental Validation
The framework was evaluated using a simulated system of diffusively coupled oscillators with three nodes. The system switches between two distinct topologies based on its position relative to a spherical boundary in the state space: a complete graph ($K_3$) and a directed 3-cycle ($C_3$).
Key Results
The algorithm demonstrated rapid and precise convergence. As shown in the performance metrics, both the network dynamics identification error and the adjacency matrix mismatch dropped to near-zero within five iterations. Crucially, the framework successfully recovered the exact binary structure of the hidden graphs without any prior labels.
Furthermore, by applying a soft-margin Support Vector Machine (SVM) to the identified modes, the system was able to reconstruct the hidden switching surface. The identified boundary closely matched the ground-truth sphere, demonstrating that the framework captures not just the "what" of the network changes, but also the "why."
Conclusion and Future Work
This research provides a robust, label-free method for identifying switching network systems through convex relaxation. By explicitly modeling the system as a network of interacting components rather than a monolithic black box, the approach offers high interpretability. Future work will focus on scaling these semidefinite relaxations for high-dimensional networks and exploring more complex interaction models, such as hypergraphs, to capture multi-node dependencies.
Summary updated