Application of Langevin Dynamics to Advance the Quantum Natural Gradient Optimization Algorithm

Technical Summary: Application of Langevin Dynamics to Advance the Quantum Natural Gradient Optimization Algorithm

Problem Statement
Optimization of variational quantum circuits (VQCs) in hybrid quantum-classical algorithms, such as the Variational Quantum Eigensolver (VQE) and Quantum Approximate Optimization Algorithm (QAOA), often suffers from the "overparametrization" problem. While the Quantum Natural Gradient (QNG) algorithm was proposed to mitigate this by utilizing the Quantum Geometric Tensor (QGT) to define a reparametrization-invariant descent direction, the basic QNG algorithm frequently struggles with non-convex loss landscapes. Specifically, it tends to become trapped in local minima, saddle points, and plateaus, limiting its effectiveness in complex optimization tasks.

Methodology
The authors propose a novel optimization algorithm, Momentum-QNG, derived by applying Langevin dynamics to the QNG framework.

• Theoretical Derivation: The study begins by modeling the optimization process as the discrete-time solution of the Langevin equation. By introducing a stochastic force term based on the QNG gradient ($\hat{f} = -g^{-1}(\theta)\nabla L(\theta)$) and incorporating a viscous friction term, the authors derive a discrete update rule.
• Algorithm Formulation: The resulting update rule is:
  $$\Delta\theta_{n+1} = \rho\Delta\theta_n - \eta \cdot g^{-1}(\theta_n)\nabla L(\theta_n)$$
  Here, $\rho$ acts as a momentum coefficient derived from the friction coefficient, and $\eta$ is the learning rate. This formulation generalizes the basic QNG (which is recovered when $\rho=0$) by adding an inertial term.
• Mechanism: The momentum term allows the optimizer to accumulate velocity, enabling it to escape local minima and plateaus where local gradients vanish. Theoretical analysis (Eq. 17) suggests that the mean jump length of the variational parameters near an extremum increases with the momentum coefficient, facilitating broader exploration of the parameter space.
• Implementation: The algorithm was implemented within the PennyLane quantum computing package. The Quantum Geometric Tensor was computed using a block-diagonal approximation with a regularization coefficient $\lambda = 0.5$.

Benchmarking and Results
The Momentum-QNG algorithm was benchmarked against the basic QNG, standard Momentum, and Adam optimizers across three distinct variational quantum problems:

1. Investment Portfolio Optimization (Ising Spin-Glass Model):

   • Setup: VQE optimization for $N=6, 11, 12$ qubits.
   • Results: Momentum-amended optimizers (Momentum-QNG, Momentum, Adam) significantly outperformed the basic QNG in terms of energy convergence ($\Delta E$). While the basic QNG showed the fastest convergence speed, it resulted in the highest energy misfit. Adam demonstrated the most robust performance with the widest convergence domain.

2. Quantum Sherrington-Kirkpatrick (SK) Model:

   • Setup: Ground state energy minimization for $N=8$ qubits under varying transverse field strengths ($g$).
   • Results: Performance varied by regime:
     • For strong fields ($g=0.1$), the basic QNG achieved the lowest error, consistent with previous findings.
     • For intermediate fields ($g=10^{-3}$), Momentum and Momentum-QNG performed equally well.
     • For weak fields ($g=10^{-5}$), representing a strong spin-glass regime with pronounced local minima, Momentum-QNG outperformed all other algorithms.
   • Observation: The authors attribute the superior performance of Momentum-QNG in the strong spin-glass regime to a synergistic effect between the quantum geometric tensor and the momentum term, allowing the algorithm to navigate complex energy landscapes more effectively.

3. Minimum Vertex Cover (QAOA):

   • Setup: Optimization on graphs with $N=4$ and $N=8$ vertices.
   • Results: Momentum-QNG, Momentum, and Adam achieved similar maximal quality ratios, significantly outperforming the basic QNG. Adam maintained the widest convergence domain, while Momentum-QNG performed comparably to Adam within its domain.

Key Contributions

• Derivation of Momentum-QNG: The paper establishes a theoretical link between Langevin dynamics and quantum optimization, deriving a generalized QNG algorithm that incorporates momentum.
• Performance Validation: Through extensive numerical benchmarks, the study demonstrates that adding a momentum term to QNG improves the ability to escape local minima and plateaus, particularly in non-convex, overparametrized landscapes.
• Open Source: The authors provide open-source code for the Momentum-QNG algorithm to facilitate further research.

Significance and Claims
The paper claims that Momentum-QNG represents a significant advancement for optimizing variational quantum circuits, specifically in scenarios where the loss landscape is rugged and prone to local minima. The authors emphasize that while the basic QNG is effective for convex or simpler problems, the addition of the momentum term is crucial for navigating the complex, non-convex landscapes typical of strong spin-glass regimes and other difficult optimization tasks. The study concludes that the combination of the quantum geometric tensor (for reparametrization invariance) and momentum (for inertia) creates a synergistic effect that enhances optimization performance in challenging quantum regimes. The authors remain modest, noting that Adam often remains the most robust overall, but highlighting that Momentum-QNG offers a distinct advantage in specific high-difficulty regimes where other momentum-based methods may not fully leverage the quantum geometric structure.