Quantum Pontryagin Neural Networks in Gamkrelidze Form Subjected to the Purity of Quantum Channels

Nahid Binandeh Dehaghani, A. Pedro Aguiar, Rafał WiśniewskiIEEE Control. Syst. Lett 2023

Abstract

We investigate a time and energy minimization optimal control problem for open quantum systems, whose dynamics is governed through the Lindblad (or Gorini-Kossakowski-Sudarshan-Lindblad) master equation. The dissipation is Markovian time-independent, and the control is governed by the Hamiltonian of a quantum-mechanical system. We are specifically interested to study the purity in a dissipative system constrained by state and control inputs. The idea for solving this problem is by the combination of two following techniques. We deal with the state constraints through Gamkrelidze revisited method, while handling control constraints through the idea of saturation functions and system extensions. This is the first time that quantum purity conservation is formulated in such framework. We obtain the necessary conditions of optimality through the Pontryagin Minimum Principle. Finally, the resulted boundary value problem is solved by a Physics-Informed Neural Network (PINN) approach. The exploited Pontryagin PINN technique is also new in quantum control context. We show that these PINNs play an effective role in learning optimal control actions.

Introduction

In the development of functional quantum computers, environmental noise remains a primary obstacle. Most quantum systems are "open," interacting with their surroundings in ways that lead to dissipation and decoherence. This interaction causes a "pure" quantum state to become "mixed," resulting in the loss of quantum information—a process quantified by a decrease in state purity.

To mitigate these effects, researchers employ Optimal Control Theory (OCT) to manipulate systems efficiently before environmental noise degrades the state. However, this task involves a complex "triple threat": balancing the trade-off between speed and energy, adhering to strict state constraints (maintaining purity), and overcoming the mathematical difficulty of solving the resulting Boundary Value Problems (BVPs).

Methodology

This research introduces a multi-layered mathematical framework to address these challenges:

1. System Dynamics and Constraints

The system is modeled using the Lindblad master equation, the standard for Markovian dissipative systems. To handle constraints, the authors employ two distinct methods:

  • State Constraints (Purity): The Gamkrelidze revisited method is used to incorporate purity bounds directly into the optimality conditions. This ensures the state remains sufficiently pure throughout the entire control duration, not just at the final time.
  • Control Constraints: To respect hardware limits, saturation functions and system extensions are used to map unconstrained variables into physically permissible ranges without the need for hard-clipping.

2. Pontryagin Neural Networks (PoNNs)

The core innovation lies in the use of Pontryagin Neural Networks (PoNNs), a specialized type of Physics-Informed Neural Network (PINN) based on the Theory of Functional Connection (TFC). Unlike standard PINNs, TFC-based PoNNs utilize "constrained expressions" that analytically satisfy boundary conditions. This allows the network to focus its learning capacity entirely on minimizing the physics-based loss derived from the Pontryagin Minimum Principle (PMP).

Numerical Validation and Results

The framework was validated through numerical simulations of a two-level atom interacting with a vacuum environment. This benchmark problem demonstrates the PoNN's ability to learn optimal control paths in a dissipative qubit system.

Population dynamics showing (a) Rabi oscillations, (b) decay, (c) damped oscillations, and (d) the optimized state transfer.

The simulation results confirmed that the PoNN could identify optimal control actions to transition the system from an excited state to a ground state while actively resisting environmental decay. As shown in Figure 2(d), the optimized transfer achieves a complete population swap within the allotted time frame.

Fidelity evolution comparing purity-constrained vs. unconstrained control paths.

A key finding of the study is the quantifiable trade-off between purity preservation and fidelity. As illustrated in Figure 1, imposing strict purity constraints (the green line) results in a slower rise in fidelity compared to the unconstrained case (the yellow line). However, the constrained approach ensures the system remains within the required purity threshold throughout the process, which is critical for maintaining the integrity of quantum information.

Conclusion

This work demonstrates that turning a calculus-of-variations problem into a machine-learning optimization problem via PoNNs offers a robust alternative to traditional numerical solvers like "shooting methods." By embedding boundary conditions into the network architecture, the researchers achieved faster convergence and more accurate adherence to physical constraints.

Future work will focus on scaling this approach to multi-qubit systems, addressing the "curse of dimensionality," and exploring non-Markovian environments where system history influences future dissipation.

Summary updated

Abstract

We investigate a time and energy minimization optimal control problem for open quantum systems, whose dynamics is governed through the Lindblad (or Gorini-Kossakowski-Sudarshan-Lindblad) master equation. The dissipation is Markovian time-independent, and the control is governed by the Hamiltonian of a quantum-mechanical system. We are specifically interested to study the purity in a dissipative system constrained by state and control inputs. The idea for solving this problem is by the combination of two following techniques. We deal with the state constraints through Gamkrelidze revisited method, while handling control constraints through the idea of saturation functions and system extensions. This is the first time that quantum purity conservation is formulated in such framework. We obtain the necessary conditions of optimality through the Pontryagin Minimum Principle. Finally, the resulted boundary value problem is solved by a Physics-Informed Neural Network (PINN) approach. The exploited Pontryagin PINN technique is also new in quantum control context. We show that these PINNs play an effective role in learning optimal control actions.
BibTeX
@article{dehaghani2023quantum,
  author = {Nahid Binandeh Dehaghani and A. Pedro Aguiar and Rafał Wiśniewski},
  title = {{Quantum Pontryagin Neural Networks in Gamkrelidze Form Subjected to the Purity of Quantum Channels}},
  journal = {IEEE Control. Syst. Lett.},
  year = {2023},
}
Figure