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

A Quantum Natural Gradient (QNG) algorithm for optimization of variational quantum circuits has been proposed recently. In this study, we employ the Langevin equation with a QNG stochastic force to demonstrate that its discrete-time solution gives a generalized form of the above-specified algorithm, which we call Momentum-QNG. Similar to other optimization algorithms with the momentum term, such as the Stochastic Gradient Descent with momentum, RMSProp with momentum and Adam, Momentum-QNG is more effective to escape local minima and plateaus in the variational parameter space and, therefore, demonstrates an improved performance compared to the basic QNG. In this paper we benchmark Momentum-QNG together with the basic QNG, Adam and Momentum optimizers and explore its convergence behaviour. Among the benchmarking problems studied, the best result is obtained for the quantum Sherrington-Kirkpatrick model in the strong spin glass regime. Our open-source code is available at https://github.com/borbysh/Momentum-QNG

💡 Research Summary

The paper introduces a novel optimizer for variational quantum circuits called Momentum‑QNG, which augments the previously proposed Quantum Natural Gradient (QNG) method with a momentum term derived from Langevin dynamics. QNG rescales the gradient by the inverse of the Fubini‑Study metric tensor, thereby mitigating over‑parameterization and providing a re‑parameterization‑invariant descent direction. However, QNG alone often gets trapped in local minima, saddles, or plateaus when the loss landscape is non‑convex.

To address this, the authors start from the multidimensional Langevin equation, adding a viscous friction term (γ) and a stochastic white‑noise term (R(t)) to Newton’s second law. By discretizing the continuous‑time equation, they obtain an update rule that matches the classic stochastic gradient descent with momentum (SGD‑Momentum). In this formulation, the momentum coefficient ρ and learning rate η are functions of the friction γ and the time step Δt. The relationship between the noise variance, temperature, and momentum shows that increasing ρ enlarges the typical parameter jump size, facilitating escape from shallow basins.

Substituting the deterministic QNG force f = −g⁻¹∇L into the Langevin force yields the discrete‑time update
Δθₙ₊₁ = ρ Δθₙ − η g⁻¹(θₙ)∇L(θₙ),
which reduces to standard QNG when ρ = 0. This Momentum‑QNG therefore combines the geometry‑aware scaling of QNG with the inertial exploration of momentum‑based optimizers.

The authors implement four optimizers—Adam, Momentum (SGD‑Momentum), basic QNG, and Momentum‑QNG—within the PennyLane framework, using a block‑diagonal approximation of the metric tensor and a regularization λ = 0.5. All experiments start from zero initial parameter updates and share identical random seeds for fair comparison.

Benchmark 1: Portfolio optimization mapped to an N‑spin Ising model (N = 6, 11, 12). Across 200 trials per setting, Momentum‑QNG, Momentum, and Adam achieve significantly lower energy errors and more stable convergence than QNG, with Adam showing the broadest learning‑rate tolerance.

Benchmark 2: Quantum Sherrington‑Kirkpatrick (SK) spin‑glass model (N = 8) under a transverse field g. For a strong field (g = 0.1) the original QNG attains the smallest ground‑state error, confirming its strength in highly rugged landscapes. For weaker fields (g = 10⁻³ and 10⁻⁵), Momentum‑QNG matches or surpasses the other methods, demonstrating that momentum aids exploration when the effective temperature is low.

Benchmark 3: Quantum Approximate Optimization Algorithm (QAOA) applied to the Minimum Vertex Cover problem (graphs with 4 and 8 vertices). Momentum‑QNG and Adam reach comparable maximal quality ratios, outperforming QNG. Momentum‑QNG shows a moderate convergence domain, while Adam enjoys the widest.

Overall, the empirical results validate the theoretical claim that adding a momentum term to QNG enlarges the stochastic “jump” size (as per Eq. 17) and improves the ability to traverse flat or deceptive regions of the loss surface. The paper concludes that Langevin‑derived Momentum‑QNG offers a simple yet effective enhancement to quantum circuit optimization, especially for non‑convex problems. Future directions suggested include adaptive schedules for temperature and friction, more accurate metric‑tensor inversions, and scaling to larger qubit counts.