arXiv:2610.30301
Non-adiabatic Optimisation via Localisation-Delocalisation Transitions
Ryan Smith, Sabrina Yue Wang, Filippo Maria Gambetta, Raul A. Santos
Abstract
The promise of efficiently solving classical optimisation problems on quantum computers is tantalizing. As many such problems can be mapped to approximating the ground state of spin-glasses, paradigmatic models for studying many-body localisation, it is natural to ask if some of the intuition coming from quantum statistical mechanics can help guide the construction of novel quantum optimisation algorithms. In this work, we discuss how continuous reverse quantum annealing crossing a many-body localisation transition can be accelerated by considering the density of Landau-Zener crossings as the system is quenched. We provide an interpretation of this behaviour in terms of an effective Fokker-Planck equation for the success probability of the algorithm. This interpretation suggests an improvement in the required annealing time as a function of the system size $N$, which goes from $\mathcal{O}({\rm poly}(N))$ to $\mathcal{O}(1)$ for reference states with non-vanishing energy density. We show that the continuous dynamics can be Trotterised with a number of steps that remains modest as system size increases, making it appealing for near-term devices. For the resulting circuit implementation, we provide a quantitative strategy for determining the parameters of the time-dependent Hamiltonian. We compare the performance of our digitised algorithm, which we call FRIO, against that of the QAOA for finding the ground state of spin-glass problems, in particular the SK model. The numerical results suggest that with suitably selected circuit parameters which can be efficiently found using thermalizing dynamics, FRIO almost always outperforms the conventional QAOA ansatz with pre-optimized angles.
AI Summary
Can reverse quantum annealing exploit localisation–delocalisation transitions to improve classical spin-glass optimisation without adiabatic evolution? This paper introduces Fast Reverse Iterative Optimisation (FRIO), which starts from a classical reference bitstring, applies reference-aligned longitudinal and transverse fields, then quenches both fields off. Computational-basis measurements replace the reference only when they lower its energy. The protocol is also digitised using first- and second-order Trotter decompositions. A coarse-grained model treats avoided crossings as independent Landau–Zener events and derives an effective Fokker–Planck description. Assuming exponentially many available crossings, polynomially scaling gaps and a favourable downward-transition asymmetry, it predicts that an $\mathcal{O}(1)$ annealing time per cycle can yield an energy-improvement probability above $1/2$. This is not a bound on total ground-state-finding time. A parameter-selection procedure uses short-time thermalisation of a four-qubit marginal’s inverse participation ratio. Exact-diagonalisation and circuit simulations study SK instances and weighted degree-$3,4,5$ regular graphs, with optimisation benchmarks reaching $N=18$. At $B_z=1$ and $B_x/B_z\approx1$, digitised FRIO generally produces more ground-state samples than QAOA with ensemble-based or instance-optimised angles, using comparable circuit resources and equal measurement budgets. Benchmarks use ten instances per size and $50{,}000$ measurements per method. The stated limitations include unverified large-system parameter transfer, declining improvement probabilities near the spectrum bottom, and assumptions of independent crossings and slowly varying drift and diffusion. One-step Trotterisation deteriorates at larger simulated sizes; larger systems may require longer quenches and more Trotter steps.
\Cref{subsec:nonadiabatic} was partially developed through discussions with Claude. Claude was also used to identify typographical errors throughout the manuscript.
Central claim may not hold: Non-vanishing reference energy density does not establish the positive, suitably size-dependent asymmetry eta required by the paper's constant-time success argument. The conditional stochastic analysis is not validated in the main numerical operating regime, and simulations of at most 18 spins do not establish the advertised asymptotic annealing-time improvement. The exponential crossing rate, polynomial gaps and independent constant-coefficient jump model predict exponentially growing fixed-time energy variance, contradicting the polynomial bandwidth of the final SK Hamiltonian and invalidating the central asymptotic extrapolation. The displayed Gaussian integrates below the reference energy to one-half erfc(x), not one-half erfc(-x), and the assigned drift sign is opposite to the stated downward-transition bias. Step 1's claimed bias threshold is generally incorrect because the required field depends on every competitor's energy difference divided by twice its Hamming distance from the reference.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 25 pages, 5 figures + supplementary information