Quantum Computing Achieves up to 10% Improvement with Novel LOTUS Optimisation Schedules

Key Takeaways

• LOTUS reduces QAOA parameter space to constant dimensionality
• Achieves up to 27% higher expectation values vs L‑BFGS‑B
• Cuts optimisation iterations by over 90% compared to Powell
• Enables depth transferability, initializing deeper circuits efficiently
• Demonstrated on MaxCut; future work targets broader NP‑hard problems

Summary

Researchers at Viettel High Technology Industries have introduced LOTUS, a Layer‑Ordered Temporally Unified Schedule framework that re‑parameterises QAOA circuits via a Hybrid Fourier‑Autoregressive mapping. By collapsing the high‑dimensional parameter space to a constant‑size representation, LOTUS delivers up to 27.2 % higher expectation values than leading classical optimisers such as L‑BFGS‑B and reduces the number of optimisation iterations by more than 90 %. The method also shows depth transferability, allowing schedules trained on shallow circuits to initialise deeper ones, and is validated on the MaxCut problem. The authors suggest extending the approach to other NP‑hard tasks like MaxSAT and TSP.

Pulse Analysis

Quantum Approximate Optimisation Algorithms have emerged as a leading candidate for solving combinatorial problems on near‑term quantum processors, yet their practical deployment is hampered by the exponential growth of tunable parameters as circuit depth increases. Traditional gradient‑based or derivative‑free optimisers often stall in noisy landscapes, leading to sub‑optimal solutions and prohibitive runtimes. The newly proposed LOTUS framework tackles this bottleneck by recasting the entire schedule into a low‑dimensional dynamical system, thereby sidestepping the curse of dimensionality that plagues conventional approaches.

At the heart of LOTUS lies a Hybrid Fourier‑Autoregressive (HFA) mapping that replaces independent layer‑wise angles with a unified set of Fourier coefficients and autoregressive terms. This representation enforces global temporal coherence across the circuit while preserving enough local flexibility to adapt to problem‑specific features. Empirical tests on the MaxCut benchmark reveal up to 27.2 % higher expectation values compared with L‑BFGS‑B and a 20.8 % edge over COBYLA, while requiring more than 90 % fewer optimisation iterations than Powell, SLSQP, or TNC. Such gains translate directly into reduced quantum‑hardware runtime and lower classical preprocessing costs.

The ability to collapse the optimisation space to O(1) complexity opens a clear pathway for scaling QAOA to utility‑scale problems such as logistics, finance, and materials design. By reusing schedules trained on shallow circuits as initial seeds for deeper layers, LOTUS further accelerates convergence, making it attractive for cloud‑based quantum services that charge per optimisation minute. Although the current study focuses on MaxCut, the authors plan to extend the HFA ansatz to other NP‑hard formulations like MaxSAT, travelling‑salesperson, and generic QUBO models, promising broader industry impact as quantum hardware matures.