Optimisation Schedules Transfer Between Quantum Problem Sizes

Scientists are tackling the challenge of scaling the Quantum Approximate Optimisation Algorithm (QAOA) for complex combinatorial problems. Ugo Nzongani from Aix-Marseille Universit e, Universit e de Toulon, CNRS, LIS, working with Dylan Laplace Mermoud from UMA, ENSTA, Institut Polytechnique de Paris, and Arthur Braida from Universit e Paris Cit e, CNRS, IRIF, demonstrate a novel schedule-learning framework that transfers optimised adiabatic control strategies from smaller quantum circuits to larger, more complex systems. This collaborative research exploits the observed concentration and transferability of optimal QAOA angles, significantly reducing the number of parameters requiring classical optimisation from a function of circuit depth to just two global hyperparameters. By extracting spectral gap profiles from small problems and constructing a continuous schedule, the team achieves competitive approximation ratios on random QUBO and 3-regular MaxCut instances, offering a parameter-efficient and scalable strategy to mitigate the limitations of current QAOA implementations and the associated barren plateau phenomena.

Scientists are edging closer to unlocking the potential of quantum computing with a technique to dramatically simplify complex calculations. The method overcomes a key hurdle in scaling up these systems, promising more powerful optimisation for real-world problems. The core challenge hindering wider application of QAOA lies in the exponentially increasing difficulty of optimising the numerous parameters required as problem size grows.

This work circumvents that limitation by transferring knowledge gleaned from smaller systems to larger ones, dramatically reducing the number of parameters needing classical optimisation. Researchers demonstrate that optimal QAOA configurations exhibit predictable patterns across different problem sizes, allowing them to construct efficient schedules for quantum computations.

The study focuses on leveraging the spectral gap, the difference between the ground state and first excited state of the problem Hamiltonian, as a guiding principle for designing these schedules. By extracting the spectral gap profile from small-scale instances, the team created a continuous schedule governed by a simple equation relating the rate of change of the schedule to the instantaneous gap.

Discretizing this schedule yields a closed-form expression for all QAOA angles, compressing the classical optimisation task from a number of parameters dependent on circuit depth to two global hyperparameters. This parameter compression not only reduces computational overhead but also mitigates the risk of encountering barren plateaus, a phenomenon where gradients vanish, stalling the optimisation process.

This innovative approach effectively maps the continuous process of adiabatic evolution, slowly transitioning between simple and complex Hamiltonians, onto the discrete steps of a QAOA circuit. Numerical simulations on both random Quadratic Unconstrained Binary Optimisation problems and 3-regular MaxCut instances confirm the effectiveness of this transfer learning technique.

The learnt schedules transfer well to larger systems, achieving competitive approximation ratios without the need for extensive re-optimisation. Gap-informed schedule transfers offer a scalable and parameter-efficient strategy for harnessing the power of QAOA in tackling increasingly complex optimisation challenges. This advancement could unlock new possibilities for applying quantum computing to real-world problems in fields like logistics, finance, and materials science.

Spectral gap profile extraction simplifies QAOA angle optimisation and parameter count

Learnt schedules yielded closed-form expressions for all QAOA angles, compressing the classical optimisation task from 2p parameters to only 2, irrespective of circuit depth. This parameter reduction substantially mitigates classical optimisation overhead and lessens sensitivity to barren plateau phenomena. Numerical simulations performed on both random QUBO and 3-regular MaxCut instances demonstrated effective transfer of the learnt schedules to larger systems, achieving competitive approximation ratios.

Specifically, the research focused on extracting the spectral gap profile from small problems and constructing a continuous schedule governed by the equation s(t) = α/ (β + t), where s(t) represents the instantaneous gap and α and β are global hyperparameters. Discretization of this schedule resulted in explicit formulas for all QAOA angles, a significant simplification of the optimisation process.

The study assessed performance on random QUBO instances with up to 12 qubits and 3-regular MaxCut graphs containing up to 16 nodes. Across these instances, the learnt schedules consistently achieved approximation ratios comparable to those obtained through full classical optimisation, despite the drastic reduction in trainable parameters. The parameter compression, from 2p to 2, represents a substantial computational advantage, particularly for deeper circuits where classical optimisation becomes intractable.

Furthermore, the research highlighted the robustness of the approach, demonstrating that the optimised hyperparameters α and β transfer effectively across different problem sizes. The spectral gap profiles extracted from smaller instances accurately predicted the optimal annealing schedule for larger systems, indicating a degree of universality in the learnt control strategies.

This transferability is crucial for scaling QAOA to tackle more complex optimisation problems. The method’s ability to circumvent the limitations imposed by barren plateaus suggests a pathway towards developing more reliable and efficient quantum optimisation algorithms.

Spectral gap guided schedule learning for adiabatic quantum optimisation

A continuous schedule-learning framework underpinned this work, transferring spectral-gap-informed adiabatic control strategies from small-scale problem instances to larger systems. Initially, the spectral gap profile was extracted from smaller problems, characterising how the minimum energy difference between the ground state and the first excited state changes with system parameters.

This profile then informed the construction of a continuous schedule, mathematically defined by the equation ∂ts = κg(s), where ‘ts’ represents time, ‘κ’ and ‘q’ are global hyperparameters controlling the schedule’s shape, and ‘g(s)’ denotes the instantaneous spectral gap. The choice of focusing on the spectral gap stems from its direct relationship to the adiabatic theorem, ensuring a smooth transition to the problem’s ground state and minimising errors during computation.

Discretization of this continuous schedule was a key methodological innovation, yielding closed-form expressions for all QAOA angles. This process effectively transformed the classical optimisation task, reducing the number of parameters to optimise from 2p, where ‘p’ represents the number of layers in the quantum circuit, to 2, independent of circuit depth.

This drastic parameter compression is crucial for mitigating the classical computational overhead associated with training deep quantum circuits and for reducing sensitivity to barren plateau phenomena, which hinder optimisation in high-dimensional parameter spaces. The resulting angles are then directly used to define the unitary operations within the QAOA circuit.

To assess the transferability and efficacy of the learnt schedules, numerical simulations were performed on both random QUBO and 3-regular MaxCut instances. QUBO (Quadratic Unconstrained Binary Optimisation) problems represent a versatile class of combinatorial optimisation challenges, while 3-regular MaxCut instances provide a structured test case with well-defined properties.

These simulations allowed for a comparative analysis of the approximation ratios achieved with the transferred schedules against those obtained through conventional optimisation methods, demonstrating the scalability and parameter efficiency of the proposed strategy for QAOA. The selection of these instance types aimed to provide a robust evaluation across diverse problem landscapes.

Transfer learning of QAOA parameters alleviates classical computational demands

The relentless pursuit of scalable quantum algorithms has long been hampered by a frustrating paradox: the very techniques designed to harness quantum mechanics often succumb to classical bottlenecks. By cleverly leveraging the predictable behaviour of optimal QAOA parameters, their tendency to concentrate and transfer across problem sizes, researchers have dramatically reduced the computational burden of finding those parameters.

The core idea, transferring insights from smaller, solvable instances to larger, intractable ones, feels intuitively elegant. What distinguishes this approach isn’t necessarily a leap in quantum performance, but a significant reduction in the classical pre-processing required. For QAOA to function, parameters must be optimised classically, a task that rapidly becomes overwhelming as problem complexity increases and is prone to the ‘barren plateau’ phenomenon where gradients vanish.

Compressing the parameter space from potentially thousands to a handful represents a genuine advance, potentially unlocking QAOA’s potential on near-term hardware. However, the transferability of these ‘learnt schedules’ isn’t guaranteed to hold universally. The simulations focused on specific problem types, random QUBO and 3-regular MaxCut, and the extent to which these findings generalise to real-world optimisation challenges remains an open question.

Moreover, while parameter compression is valuable, it doesn’t address the fundamental limitations of QAOA itself, such as its inability to demonstrably outperform classical algorithms on many problems. Future work will likely explore adaptive transfer strategies, tailoring the schedule learning process to the specific characteristics of each problem instance, and combining this approach with other techniques for mitigating barren plateaus. The broader goal remains clear: to bridge the gap between theoretical promise and practical quantum advantage.