Steering the dynamics by controlling the temporal interaction network

Technical Summary: Steering the Dynamics by Controlling the Temporal Interaction Network

Problem Statement
Real-world coupled dynamical systems often possess interaction structures and strengths that evolve or adapt over time, a feature poorly captured by static network models. While traditional control strategies typically rely on additive control inputs (directly influencing node states), controlling the interaction network itself presents a more challenging problem due to its multiplicative nature. The paper addresses the question of how to steer the state of a dynamical system to desired trajectories by tuning its temporal interaction network (i.e., the time-dependent coupling weights) rather than applying external forces to the nodes.

Methodology
The authors propose a framework based on nonlinear optimal control utilizing the adjoint method. The approach is structured as follows:

1. System Formulation: The system is modeled as a network of $N$ nodes with states $x_i$. The dynamics are governed by internal flows and a coupling term weighted by a time-dependent matrix $W_{ij}(t)$. The control input $U(t)$ is identified with the coupling matrix $W(t)$.
2. Optimization Objective: A performance metric (cost function) $C(x, U)$ is defined, comprising two components:
   • Task Error ($C_e$): Quantifies the deviation of node states from prescribed target trajectories.
   • Control Effort ($C_u$): Quantifies the magnitude of the control input (coupling weights) applied.
3. Adjoint Method Derivation: To optimize the control signal $U(t)$, the authors define a Lagrangian function incorporating the system dynamics as constraints. They derive the gradient of this Lagrangian with respect to the control inputs. This involves:
   • Solving the forward state dynamics from an initial condition.
   • Solving the adjoint state evolution backward in time from a terminal condition ($\lambda(T)=0$).
   • Computing the gradient using the product of the adjoint state and the state variables.
4. Specific Application: The general framework is applied to a linear time-variant system where internal dynamics drive nodes toward natural states and coupling drives them toward neighbors. The gradient expression is explicitly derived for this linear case, allowing for numerical optimization via gradient descent (specifically using the Adam optimizer).

Key Contributions

• General Control Framework: The paper establishes a theoretical framework for controlling dynamical systems by manipulating the connectivity matrix (multiplicative control) rather than node states (additive control).
• Gradient Derivation: It provides a rigorous derivation of the Lagrangian gradient using the adjoint method, enabling the use of gradient-based optimization algorithms to find optimal time-dependent coupling weights.
• Constraint Flexibility: The framework allows for various constraints on the control matrix, such as restricting control to off-diagonal terms (inter-node coupling only) or sparsifying the control to a fraction of possible links.

Results
The framework was numerically tested on a network of $N=50$ nodes with linear time-variant dynamics. Two primary scenarios were explored:

1. Steering to a Common Trajectory: Nodes with heterogeneous natural states were steered to follow a single, complex target signal.

   • Full Control: When all off-diagonal coupling weights were controllable, the system successfully reached the target.
   • Sparse Control: The system remained effective even when control was restricted to only 8% of the total possible pairwise interactions.
   • Cost Analysis: Reducing the number of controlled edges increased both the precision cost (deviation from target) and the control cost (energy required), as expected. However, the system still converged to the target trajectory.
   • Intrinsic State Distribution: Systems with natural states closer to the target required less control effort than those with broadly distributed natural states.

2. Steering to Distinct Trajectories: The framework successfully guided nodes with similar intrinsic states to three distinct, amplitude-varying target patterns simultaneously, again using control over only 8% of the network links.

Significance and Claims
The paper claims that the presented workflow offers a viable method to steer the dynamics of systems with artificial or engineered interactions that are tunable. Key points of significance include:

• Efficiency: The adjoint method allows for the efficient computation of gradients for high-dimensional systems, making the optimization of time-varying coupling networks computationally feasible.
• Sparse Control: The results demonstrate that full control over the network is not necessary; steering can be achieved by manipulating a small fraction (e.g., 8%) of the interaction links.
• Applicability: While the numerical examples use a simple linear model, the authors state the framework is general and can be leveraged for more complex dynamical systems, including artificially coupled neuronal models and engineered systems (e.g., vehicular platoons or swarming robots).
• Future Outlook: The authors suggest that future work should extend this framework to include application-specific constraints (e.g., in neuroscience where connectivity emulation has physical limits) and explore combining nodal control with interaction control.

The paper maintains a modest tone, acknowledging that the optimization problem is typically non-convex (potentially yielding local minima) and that the current numerical illustrations rely on a simplified linear model, though the underlying theory is general.